aLoRA retrain of genai-attack-vector on granite-4.0-350m (gated on assistant invocation tokens)

adapter_config.json

@@ -32,23 +32,21 @@
   "qalora_group_size": 16,
   "r": 32,
   "rank_pattern": {
-    "down_proj": 32,
-    "gate_proj": 32,
+    "input_linear": 32,
     "k_proj": 32,
     "o_proj": 32,
+    "output_linear": 32,
     "q_proj": 32,
-    "up_proj": 32,
     "v_proj": 32
   },
   "revision": null,
   "target_modules": [
-    "up_proj",
     "o_proj",
-    "q_proj",
-    "gate_proj",
-    "down_proj",
     "v_proj",
-    "k_proj"
+    "k_proj",
+    "output_linear",
+    "q_proj",
+    "input_linear"
   ],
   "target_parameters": null,
   "task_type": "CAUSAL_LM",

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ecd2ea37998d533a6a40fa093b8f1945d89505a7c42b4d282dfbfd7a4f0a27b4
-size 56936568
+oid sha256:9e5bd03315e52a2233a1bfa553fa9e642ecff328206d9f84089423c4f35d8541
+size 53261144

checkpoint-40/adapter_config.json

@@ -32,23 +32,21 @@
   "qalora_group_size": 16,
   "r": 32,
   "rank_pattern": {
-    "down_proj": 32,
-    "gate_proj": 32,
+    "input_linear": 32,
     "k_proj": 32,
     "o_proj": 32,
+    "output_linear": 32,
     "q_proj": 32,
-    "up_proj": 32,
     "v_proj": 32
   },
   "revision": null,
   "target_modules": [
-    "up_proj",
     "o_proj",
-    "q_proj",
-    "gate_proj",
-    "down_proj",
     "v_proj",
-    "k_proj"
+    "k_proj",
+    "output_linear",
+    "q_proj",
+    "input_linear"
   ],
   "target_parameters": null,
   "task_type": "CAUSAL_LM",

checkpoint-40/adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e0d28e7fda8b0f8b4f0aff9d45a0b44dfe6e663b1d7fb6897889d31cce7676b9
-size 56936568
+oid sha256:17d53137d203df250e31d1f1ac13f164a811612f8fb25dc045e5e1ca22dc0a65
+size 53261144

checkpoint-40/optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ee4ebdf4710b6ed7f48615c9d51e7e8b208efac0f2a01d9f07f273f8304775eb
-size 114100074
+oid sha256:814e23b2e2f94fbce445c63403e6d29164d341a36cd5314ff0d94e5467666cb9
+size 106712578

checkpoint-40/rng_state_0.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:156a2c3d2e72cbd3a562437f1517820a9bd54ca6f55636975080785c11eb3269
+oid sha256:5c2af8a1828d94701339c8c8ab8493a447a2dbbc7e23dc2144e1925045b6f702
 size 15024

checkpoint-40/rng_state_1.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2f1b464d4137d3e65c50bc15cfdc2f7f74c37fdc4b19893c31c0bd4e2dd263aa
+oid sha256:a04f11a00aa55a4bc5dbe22ca2ff9fa147586d86af752c751f80caba335bd73e
 size 15024

checkpoint-40/rng_state_2.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e1e4a8a3ed7a4c0eaf78504275dc1aed8730dc507d29311a7b57eb9a6fb589d8
+oid sha256:6b6b1e452927fa56295ca8ea623963c46cd83371ef8008528c52c5db1d49fe3b
 size 15024

checkpoint-40/rng_state_3.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:43e5b5f568c5f46f46d04020b4f12412fb64033a3ec02aeb47b62b24fbc66f65
+oid sha256:8908c4736eba56990520abb0556be0bdba80e505ab5a51071eadc9fdde0e7eb1
 size 15024

checkpoint-40/trainer_state.json

@@ -10,402 +10,402 @@
   "is_world_process_zero": true,
   "log_history": [
     {
-      "entropy": 0.8529030084609985,
+      "entropy": 0.8637090027332306,
       "epoch": 0.02531645569620253,
-      "grad_norm": 99.2493667602539,
+      "grad_norm": 99.94273376464844,
       "learning_rate": 0.0,
-      "loss": 15.181689262390137,
+      "loss": 15.196868896484375,
       "mean_token_accuracy": 0.28961846977472305,
       "num_tokens": 7286.0,
       "step": 1
     },
     {
-      "entropy": 0.8989771902561188,
+      "entropy": 0.9094592481851578,
       "epoch": 0.05063291139240506,
-      "grad_norm": 94.88607025146484,
+      "grad_norm": 95.93968200683594,
       "learning_rate": 5e-05,
-      "loss": 14.95038890838623,
+      "loss": 14.926185607910156,
       "mean_token_accuracy": 0.27897340804338455,
       "num_tokens": 14954.0,
       "step": 2
     },
     {
-      "entropy": 1.1254765689373016,
+      "entropy": 1.1258608400821686,
       "epoch": 0.0759493670886076,
-      "grad_norm": 64.76227569580078,
+      "grad_norm": 64.25172424316406,
       "learning_rate": 0.0001,
-      "loss": 12.595176696777344,
+      "loss": 12.606666564941406,
       "mean_token_accuracy": 0.35700511932373047,
       "num_tokens": 21366.0,
       "step": 3
     },
     {
-      "entropy": 1.0708639919757843,
+      "entropy": 1.0687015354633331,
       "epoch": 0.10126582278481013,
-      "grad_norm": 36.60965347290039,
+      "grad_norm": 36.90766906738281,
       "learning_rate": 0.00015000000000000001,
-      "loss": 9.597381591796875,
-      "mean_token_accuracy": 0.5474753379821777,
+      "loss": 9.619781494140625,
+      "mean_token_accuracy": 0.540277749300003,
       "num_tokens": 28250.0,
       "step": 4
     },
     {
-      "entropy": 1.184021770954132,
+      "entropy": 1.1662852466106415,
       "epoch": 0.12658227848101267,
-      "grad_norm": 29.169525146484375,
+      "grad_norm": 29.526512145996094,
       "learning_rate": 0.0002,
-      "loss": 8.115850448608398,
+      "loss": 8.147899627685547,
       "mean_token_accuracy": 0.567672997713089,
       "num_tokens": 34914.0,
       "step": 5
     },
     {
-      "entropy": 0.7149983793497086,
+      "entropy": 0.727982223033905,
       "epoch": 0.1518987341772152,
-      "grad_norm": 13.915483474731445,
+      "grad_norm": 13.896689414978027,
       "learning_rate": 0.0001999145758387301,
-      "loss": 7.248701095581055,
-      "mean_token_accuracy": 0.5984602570533752,
+      "loss": 7.228446006774902,
+      "mean_token_accuracy": 0.6021915972232819,
       "num_tokens": 42532.0,
       "step": 6
     },
     {
-      "entropy": 1.2620182633399963,
+      "entropy": 1.246531367301941,
       "epoch": 0.17721518987341772,
-      "grad_norm": 44.453094482421875,
+      "grad_norm": 19.760786056518555,
       "learning_rate": 0.000199658449300667,
-      "loss": 6.674359321594238,
-      "mean_token_accuracy": 0.6354625821113586,
+      "loss": 6.584984302520752,
+      "mean_token_accuracy": 0.6384725868701935,
       "num_tokens": 49872.0,
       "step": 7
     },
     {
-      "entropy": 1.2574806809425354,
+      "entropy": 1.175276130437851,
       "epoch": 0.20253164556962025,
-      "grad_norm": 17.659940719604492,
+      "grad_norm": 12.119710922241211,
       "learning_rate": 0.0001992320579737045,
-      "loss": 6.544322967529297,
-      "mean_token_accuracy": 0.626764640212059,
+      "loss": 6.310988426208496,
+      "mean_token_accuracy": 0.6375854164361954,
       "num_tokens": 55639.0,
       "step": 8
     },
     {
-      "entropy": 1.1509844064712524,
+      "entropy": 1.0149184465408325,
       "epoch": 0.22784810126582278,
-      "grad_norm": 8.395023345947266,
+      "grad_norm": 7.977464199066162,
       "learning_rate": 0.00019863613034027224,
-      "loss": 6.157719135284424,
-      "mean_token_accuracy": 0.669881597161293,
+      "loss": 6.036639213562012,
+      "mean_token_accuracy": 0.666093721985817,
       "num_tokens": 61649.0,
       "step": 9
     },
     {
-      "entropy": 1.0907686650753021,
+      "entropy": 1.022455170750618,
       "epoch": 0.25316455696202533,
-      "grad_norm": 4.606855392456055,
+      "grad_norm": 7.014733791351318,
       "learning_rate": 0.00019787168453273544,
-      "loss": 5.892803192138672,
-      "mean_token_accuracy": 0.6802202314138412,
+      "loss": 5.827225685119629,
+      "mean_token_accuracy": 0.6803832501173019,
       "num_tokens": 69689.0,
       "step": 10
     },
     {
-      "entropy": 1.1551010608673096,
+      "entropy": 1.1049264669418335,
       "epoch": 0.27848101265822783,
-      "grad_norm": 14.0298490524292,
+      "grad_norm": 5.545648574829102,
       "learning_rate": 0.00019694002659393305,
-      "loss": 5.6713948249816895,
-      "mean_token_accuracy": 0.6972689032554626,
+      "loss": 5.525129795074463,
+      "mean_token_accuracy": 0.704516813158989,
       "num_tokens": 77528.0,
       "step": 11
     },
     {
-      "entropy": 0.8588106781244278,
+      "entropy": 0.8492173552513123,
       "epoch": 0.3037974683544304,
-      "grad_norm": 4.56960916519165,
+      "grad_norm": 6.59067964553833,
       "learning_rate": 0.0001958427482458253,
-      "loss": 5.428313255310059,
-      "mean_token_accuracy": 0.7234913557767868,
+      "loss": 5.324199199676514,
+      "mean_token_accuracy": 0.8436015695333481,
       "num_tokens": 85064.0,
       "step": 12
     },
     {
-      "entropy": 0.705646425485611,
+      "entropy": 0.6838414371013641,
       "epoch": 0.3291139240506329,
-      "grad_norm": 3.344228744506836,
+      "grad_norm": 3.6837315559387207,
       "learning_rate": 0.00019458172417006347,
-      "loss": 5.444645404815674,
-      "mean_token_accuracy": 0.8130310028791428,
+      "loss": 5.420000076293945,
+      "mean_token_accuracy": 0.816602423787117,
       "num_tokens": 91009.0,
       "step": 13
     },
     {
-      "entropy": 0.7050664573907852,
+      "entropy": 0.6191327422857285,
       "epoch": 0.35443037974683544,
-      "grad_norm": 3.312098979949951,
+      "grad_norm": 3.5020408630371094,
       "learning_rate": 0.0001931591088051279,
-      "loss": 5.357623100280762,
-      "mean_token_accuracy": 0.8219081908464432,
+      "loss": 5.348326683044434,
+      "mean_token_accuracy": 0.8181768357753754,
       "num_tokens": 97909.0,
       "step": 14
     },
     {
-      "entropy": 0.6056851297616959,
+      "entropy": 0.5245884731411934,
       "epoch": 0.379746835443038,
-      "grad_norm": 4.054323196411133,
+      "grad_norm": 3.6154603958129883,
       "learning_rate": 0.00019157733266550575,
-      "loss": 5.329057216644287,
-      "mean_token_accuracy": 0.8183465003967285,
+      "loss": 5.308533191680908,
+      "mean_token_accuracy": 0.8219696879386902,
       "num_tokens": 104187.0,
       "step": 15
     },
     {
-      "entropy": 0.46965425461530685,
+      "entropy": 0.4121474325656891,
       "epoch": 0.4050632911392405,
-      "grad_norm": 3.1617722511291504,
+      "grad_norm": 2.822946071624756,
       "learning_rate": 0.0001898390981891979,
-      "loss": 5.159331321716309,
-      "mean_token_accuracy": 0.8249307423830032,
+      "loss": 5.135751247406006,
+      "mean_token_accuracy": 0.8321253657341003,
       "num_tokens": 110300.0,
       "step": 16
     },
     {
-      "entropy": 0.4227886497974396,
+      "entropy": 0.392690971493721,
       "epoch": 0.43037974683544306,
-      "grad_norm": 6.028562545776367,
+      "grad_norm": 5.22876501083374,
       "learning_rate": 0.0001879473751206489,
-      "loss": 5.351903915405273,
-      "mean_token_accuracy": 0.8241183459758759,
+      "loss": 5.34346866607666,
+      "mean_token_accuracy": 0.8164242208003998,
       "num_tokens": 117157.0,
       "step": 17
     },
     {
-      "entropy": 0.3527509272098541,
+      "entropy": 0.3356513977050781,
       "epoch": 0.45569620253164556,
-      "grad_norm": 3.293024778366089,
+      "grad_norm": 3.489384412765503,
       "learning_rate": 0.00018590539543698854,
-      "loss": 4.932947158813477,
-      "mean_token_accuracy": 0.8373657613992691,
+      "loss": 4.941544532775879,
+      "mean_token_accuracy": 0.8337943404912949,
       "num_tokens": 123378.0,
       "step": 18
     },
     {
-      "entropy": 0.33450157940387726,
+      "entropy": 0.32025614380836487,
       "epoch": 0.4810126582278481,
-      "grad_norm": 1.6703517436981201,
+      "grad_norm": 2.2713074684143066,
       "learning_rate": 0.00018371664782625287,
-      "loss": 5.134032249450684,
-      "mean_token_accuracy": 0.8269297778606415,
+      "loss": 5.1571879386901855,
+      "mean_token_accuracy": 0.8307176530361176,
       "num_tokens": 129042.0,
       "step": 19
     },
     {
-      "entropy": 0.3399146646261215,
+      "entropy": 0.3084551766514778,
       "epoch": 0.5063291139240507,
-      "grad_norm": 2.0219178199768066,
+      "grad_norm": 1.9176894426345825,
       "learning_rate": 0.0001813848717270195,
-      "loss": 5.027883052825928,
-      "mean_token_accuracy": 0.8322668075561523,
+      "loss": 5.014953136444092,
+      "mean_token_accuracy": 0.8466087132692337,
       "num_tokens": 135747.0,
       "step": 20
     },
     {
-      "entropy": 0.29102808609604836,
+      "entropy": 0.26904235407710075,
       "epoch": 0.5316455696202531,
-      "grad_norm": 1.350559949874878,
+      "grad_norm": 1.4043225049972534,
       "learning_rate": 0.00017891405093963938,
-      "loss": 4.954178810119629,
-      "mean_token_accuracy": 0.8439806550741196,
+      "loss": 4.949811935424805,
+      "mean_token_accuracy": 0.8474528640508652,
       "num_tokens": 141848.0,
       "step": 21
     },
     {
-      "entropy": 0.3136157840490341,
+      "entropy": 0.2922343239188194,
       "epoch": 0.5569620253164557,
-      "grad_norm": 1.3057941198349,
+      "grad_norm": 1.2163947820663452,
       "learning_rate": 0.00017630840681998066,
-      "loss": 5.179681301116943,
-      "mean_token_accuracy": 0.8438494354486465,
+      "loss": 5.169545650482178,
+      "mean_token_accuracy": 0.8475807756185532,
       "num_tokens": 147566.0,
       "step": 22
     },
     {
-      "entropy": 0.2950733080506325,
+      "entropy": 0.2782471589744091,
       "epoch": 0.5822784810126582,
-      "grad_norm": 1.1281788349151611,
+      "grad_norm": 1.0059943199157715,
       "learning_rate": 0.00017357239106731317,
-      "loss": 5.171978950500488,
+      "loss": 5.164263725280762,
       "mean_token_accuracy": 0.8462631553411484,
       "num_tokens": 154620.0,
       "step": 23
     },
     {
-      "entropy": 0.27676206082105637,
+      "entropy": 0.2586139217019081,
       "epoch": 0.6075949367088608,
-      "grad_norm": 1.311663269996643,
+      "grad_norm": 1.22340726852417,
       "learning_rate": 0.00017071067811865476,
-      "loss": 5.077099800109863,
+      "loss": 5.076096534729004,
       "mean_token_accuracy": 0.8611445128917694,
       "num_tokens": 161701.0,
       "step": 24
     },
     {
-      "entropy": 0.30840296298265457,
+      "entropy": 0.28559913113713264,
       "epoch": 0.6329113924050633,
-      "grad_norm": 2.281773090362549,
+      "grad_norm": 2.1334292888641357,
       "learning_rate": 0.00016772815716257412,
-      "loss": 5.118645668029785,
-      "mean_token_accuracy": 0.8323634117841721,
+      "loss": 5.121291160583496,
+      "mean_token_accuracy": 0.8396630734205246,
       "num_tokens": 168435.0,
       "step": 25
     },
     {
-      "entropy": 0.29603931307792664,
+      "entropy": 0.26606928184628487,
       "epoch": 0.6582278481012658,
-      "grad_norm": 1.230368971824646,
+      "grad_norm": 0.8918235301971436,
       "learning_rate": 0.00016462992378609407,
-      "loss": 5.176095962524414,
-      "mean_token_accuracy": 0.839859664440155,
+      "loss": 5.148040771484375,
+      "mean_token_accuracy": 0.8437058180570602,
       "num_tokens": 175650.0,
       "step": 26
     },
     {
-      "entropy": 0.24447191506624222,
+      "entropy": 0.23234782740473747,
       "epoch": 0.6835443037974683,
-      "grad_norm": 1.7844265699386597,
+      "grad_norm": 1.8136335611343384,
       "learning_rate": 0.0001614212712689668,
-      "loss": 5.146295547485352,
-      "mean_token_accuracy": 0.8507861793041229,
+      "loss": 5.150457382202148,
+      "mean_token_accuracy": 0.8433783650398254,
       "num_tokens": 181490.0,
       "step": 27
     },
     {
-      "entropy": 0.23978285491466522,
+      "entropy": 0.24661532044410706,
       "epoch": 0.7088607594936709,
-      "grad_norm": 1.4895042181015015,
+      "grad_norm": 1.321523904800415,
       "learning_rate": 0.00015810768154019385,
-      "loss": 5.0867919921875,
-      "mean_token_accuracy": 0.8498348891735077,
+      "loss": 5.084752082824707,
+      "mean_token_accuracy": 0.8424803614616394,
       "num_tokens": 188582.0,
       "step": 28
     },
     {
-      "entropy": 0.1993725746870041,
+      "entropy": 0.19287967681884766,
       "epoch": 0.7341772151898734,
-      "grad_norm": 1.2189953327178955,
+      "grad_norm": 0.9798034429550171,
       "learning_rate": 0.00015469481581224272,
-      "loss": 5.131974697113037,
-      "mean_token_accuracy": 0.8698419630527496,
+      "loss": 5.147141456604004,
+      "mean_token_accuracy": 0.866110622882843,
       "num_tokens": 193783.0,
       "step": 29
     },
     {
-      "entropy": 0.24254225194454193,
+      "entropy": 0.23315827921032906,
       "epoch": 0.759493670886076,
-      "grad_norm": 3.0398507118225098,
+      "grad_norm": 3.1756558418273926,
       "learning_rate": 0.00015118850490896012,
-      "loss": 5.307463645935059,
-      "mean_token_accuracy": 0.8126509040594101,
+      "loss": 5.3330183029174805,
+      "mean_token_accuracy": 0.8204634040594101,
       "num_tokens": 200686.0,
       "step": 30
     },
     {
-      "entropy": 0.18977511674165726,
+      "entropy": 0.1938694752752781,
       "epoch": 0.7848101265822784,
-      "grad_norm": 1.7506808042526245,
+      "grad_norm": 1.9134807586669922,
       "learning_rate": 0.00014759473930370736,
-      "loss": 5.047242164611816,
-      "mean_token_accuracy": 0.8546684086322784,
+      "loss": 5.052967071533203,
+      "mean_token_accuracy": 0.8582398444414139,
       "num_tokens": 208440.0,
       "step": 31
     },
     {
-      "entropy": 0.2233816385269165,
+      "entropy": 0.22125443816184998,
       "epoch": 0.810126582278481,
-      "grad_norm": 1.0086487531661987,
+      "grad_norm": 1.5088441371917725,
       "learning_rate": 0.00014391965888473703,
-      "loss": 5.027800559997559,
-      "mean_token_accuracy": 0.8576832860708237,
+      "loss": 5.049187660217285,
+      "mean_token_accuracy": 0.8358375877141953,
       "num_tokens": 215131.0,
       "step": 32
     },
     {
-      "entropy": 0.19872410595417023,
+      "entropy": 0.20639389753341675,
       "epoch": 0.8354430379746836,
-      "grad_norm": 1.6391977071762085,
+      "grad_norm": 1.491873860359192,
       "learning_rate": 0.00014016954246529696,
-      "loss": 5.0654706954956055,
-      "mean_token_accuracy": 0.8516139537096024,
+      "loss": 5.055438041687012,
+      "mean_token_accuracy": 0.8552371561527252,
       "num_tokens": 222195.0,
       "step": 33
     },
     {
-      "entropy": 0.20351751521229744,
+      "entropy": 0.20158018916845322,
       "epoch": 0.8607594936708861,
-      "grad_norm": 1.5714291334152222,
+      "grad_norm": 1.46501624584198,
       "learning_rate": 0.00013635079705638298,
-      "loss": 5.170158386230469,
-      "mean_token_accuracy": 0.8566406965255737,
+      "loss": 5.16477108001709,
+      "mean_token_accuracy": 0.8602638840675354,
       "num_tokens": 229926.0,
       "step": 34
     },
     {
-      "entropy": 0.21749146282672882,
+      "entropy": 0.21367954835295677,
       "epoch": 0.8860759493670886,
-      "grad_norm": 1.6517112255096436,
+      "grad_norm": 1.7227668762207031,
       "learning_rate": 0.00013246994692046836,
-      "loss": 5.215228080749512,
+      "loss": 5.240202903747559,
       "mean_token_accuracy": 0.8288360238075256,
       "num_tokens": 235766.0,
       "step": 35
     },
     {
-      "entropy": 0.2011272795498371,
+      "entropy": 0.2060464285314083,
       "epoch": 0.9113924050632911,
-      "grad_norm": 1.4017918109893799,
+      "grad_norm": 1.615714192390442,
       "learning_rate": 0.00012853362242491053,
-      "loss": 5.1320648193359375,
+      "loss": 5.13973331451416,
       "mean_token_accuracy": 0.8528883457183838,
       "num_tokens": 242731.0,
       "step": 36
     },
     {
-      "entropy": 0.21461713686585426,
+      "entropy": 0.2112499214708805,
       "epoch": 0.9367088607594937,
-      "grad_norm": 1.77144193649292,
+      "grad_norm": 1.70606529712677,
       "learning_rate": 0.00012454854871407994,
-      "loss": 5.11080265045166,
-      "mean_token_accuracy": 0.8521855026483536,
+      "loss": 5.105003833770752,
+      "mean_token_accuracy": 0.8559168428182602,
       "num_tokens": 249818.0,
       "step": 37
     },
     {
-      "entropy": 0.19789887592196465,
+      "entropy": 0.1963396780192852,
       "epoch": 0.9620253164556962,
-      "grad_norm": 3.2191762924194336,
+      "grad_norm": 1.8689205646514893,
       "learning_rate": 0.00012052153421956342,
-      "loss": 5.070086479187012,
-      "mean_token_accuracy": 0.848893016576767,
+      "loss": 5.083700180053711,
+      "mean_token_accuracy": 0.8450468629598618,
       "num_tokens": 256043.0,
       "step": 38
     },
     {
-      "entropy": 0.1783522628247738,
+      "entropy": 0.17459847778081894,
       "epoch": 0.9873417721518988,
-      "grad_norm": 1.3007391691207886,
+      "grad_norm": 1.3199875354766846,
       "learning_rate": 0.00011645945902807341,
-      "loss": 5.1378865242004395,
-      "mean_token_accuracy": 0.8583788871765137,
+      "loss": 5.136896133422852,
+      "mean_token_accuracy": 0.8544726371765137,
       "num_tokens": 263628.0,
       "step": 39
     },
     {
-      "entropy": 0.21517179906368256,
+      "entropy": 0.19893981516361237,
       "epoch": 1.0,
-      "grad_norm": 2.357231378555298,
+      "grad_norm": 2.03135347366333,
       "learning_rate": 0.00011236926312693479,
-      "loss": 5.041143894195557,
-      "mean_token_accuracy": 0.8405462205410004,
+      "loss": 5.0191426277160645,
+      "mean_token_accuracy": 0.8476890921592712,
       "num_tokens": 266935.0,
       "step": 40
     },
@@ -415,9 +415,9 @@
       "eval_loss": NaN,
       "eval_mean_token_accuracy": 0.0,
       "eval_num_tokens": 266935.0,
-      "eval_runtime": 5.9781,
-      "eval_samples_per_second": 37.136,
-      "eval_steps_per_second": 9.368,
+      "eval_runtime": 5.2068,
+      "eval_samples_per_second": 42.637,
+      "eval_steps_per_second": 10.755,
       "step": 40
     }
   ],
@@ -447,7 +447,7 @@
       "attributes": {}
     }
   },
-  "total_flos": 560391459438592.0,
+  "total_flos": 558442705059840.0,
   "train_batch_size": 2,
   "trial_name": null,
   "trial_params": null

checkpoint-40/training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:da3e7916a9c3a9bdc1d65a3a86e68c6c06ab04fd280e02675e7357542392c8fa
+oid sha256:b5aae0a2160af9bfafbf734dc2e956a27d2ec4ded9101f3494b0466f6c3c55ff
 size 5496

checkpoint-80/adapter_config.json

@@ -32,23 +32,21 @@
   "qalora_group_size": 16,
   "r": 32,
   "rank_pattern": {
-    "down_proj": 32,
-    "gate_proj": 32,
+    "input_linear": 32,
     "k_proj": 32,
     "o_proj": 32,
+    "output_linear": 32,
     "q_proj": 32,
-    "up_proj": 32,
     "v_proj": 32
   },
   "revision": null,
   "target_modules": [
-    "up_proj",
     "o_proj",
-    "q_proj",
-    "gate_proj",
-    "down_proj",
     "v_proj",
-    "k_proj"
+    "k_proj",
+    "output_linear",
+    "q_proj",
+    "input_linear"
   ],
   "target_parameters": null,
   "task_type": "CAUSAL_LM",

checkpoint-80/adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ecd2ea37998d533a6a40fa093b8f1945d89505a7c42b4d282dfbfd7a4f0a27b4
-size 56936568
+oid sha256:9e5bd03315e52a2233a1bfa553fa9e642ecff328206d9f84089423c4f35d8541
+size 53261144

checkpoint-80/optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c9fa050e0dec60e638f2efb84ba7b8faab798c03bf0ba5d5a65f12f112ecd360
-size 114100074
+oid sha256:1a46903aa70b175d60f935d9332cd1023a1d3bc215dc31e328eba2c8e70b5443
+size 106712578

checkpoint-80/rng_state_0.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5dc2474f23a5a53cdabfdf34d0e28011700982aeb8610ac607384a176b69a792
+oid sha256:833245e699afcee735f6a6e1bce2163dfb712b42824e9d131549572f52d976f6
 size 15024

checkpoint-80/rng_state_1.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dac1653ac2bd7627a938bbf898819943e746d01bb0b238c150ec86728ce34b35
+oid sha256:42e0944fec97c4d6a98c5b1b86102667a18d80db8e8a38ab277aada9ceced00c
 size 15024

checkpoint-80/rng_state_2.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:73080767711217104a0efa90f813b1457b44079ab028f84c8b79268634e6d313
+oid sha256:2812ea526ae1b6acba02f22f1f18a45b18bd228531759fb598e508cfdcbf8e2d
 size 15024

checkpoint-80/rng_state_3.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f90541ce67a17c479ed8c2931a361d029cd473f76ad7e415bdfc30d832f320bd
+oid sha256:92b3791a5a76d2fb17a662d55e0fea058be28d67303ab0126d931a4f4272f42f
 size 15024

checkpoint-80/trainer_state.json

@@ -10,402 +10,402 @@
   "is_world_process_zero": true,
   "log_history": [
     {
-      "entropy": 0.8529030084609985,
+      "entropy": 0.8637090027332306,
       "epoch": 0.02531645569620253,
-      "grad_norm": 99.2493667602539,
+      "grad_norm": 99.94273376464844,
       "learning_rate": 0.0,
-      "loss": 15.181689262390137,
+      "loss": 15.196868896484375,
       "mean_token_accuracy": 0.28961846977472305,
       "num_tokens": 7286.0,
       "step": 1
     },
     {
-      "entropy": 0.8989771902561188,
+      "entropy": 0.9094592481851578,
       "epoch": 0.05063291139240506,
-      "grad_norm": 94.88607025146484,
+      "grad_norm": 95.93968200683594,
       "learning_rate": 5e-05,
-      "loss": 14.95038890838623,
+      "loss": 14.926185607910156,
       "mean_token_accuracy": 0.27897340804338455,
       "num_tokens": 14954.0,
       "step": 2
     },
     {
-      "entropy": 1.1254765689373016,
+      "entropy": 1.1258608400821686,
       "epoch": 0.0759493670886076,
-      "grad_norm": 64.76227569580078,
+      "grad_norm": 64.25172424316406,
       "learning_rate": 0.0001,
-      "loss": 12.595176696777344,
+      "loss": 12.606666564941406,
       "mean_token_accuracy": 0.35700511932373047,
       "num_tokens": 21366.0,
       "step": 3
     },
     {
-      "entropy": 1.0708639919757843,
+      "entropy": 1.0687015354633331,
       "epoch": 0.10126582278481013,
-      "grad_norm": 36.60965347290039,
+      "grad_norm": 36.90766906738281,
       "learning_rate": 0.00015000000000000001,
-      "loss": 9.597381591796875,
-      "mean_token_accuracy": 0.5474753379821777,
+      "loss": 9.619781494140625,
+      "mean_token_accuracy": 0.540277749300003,
       "num_tokens": 28250.0,
       "step": 4
     },
     {
-      "entropy": 1.184021770954132,
+      "entropy": 1.1662852466106415,
       "epoch": 0.12658227848101267,
-      "grad_norm": 29.169525146484375,
+      "grad_norm": 29.526512145996094,
       "learning_rate": 0.0002,
-      "loss": 8.115850448608398,
+      "loss": 8.147899627685547,
       "mean_token_accuracy": 0.567672997713089,
       "num_tokens": 34914.0,
       "step": 5
     },
     {
-      "entropy": 0.7149983793497086,
+      "entropy": 0.727982223033905,
       "epoch": 0.1518987341772152,
-      "grad_norm": 13.915483474731445,
+      "grad_norm": 13.896689414978027,
       "learning_rate": 0.0001999145758387301,
-      "loss": 7.248701095581055,
-      "mean_token_accuracy": 0.5984602570533752,
+      "loss": 7.228446006774902,
+      "mean_token_accuracy": 0.6021915972232819,
       "num_tokens": 42532.0,
       "step": 6
     },
     {
-      "entropy": 1.2620182633399963,
+      "entropy": 1.246531367301941,
       "epoch": 0.17721518987341772,
-      "grad_norm": 44.453094482421875,
+      "grad_norm": 19.760786056518555,
       "learning_rate": 0.000199658449300667,
-      "loss": 6.674359321594238,
-      "mean_token_accuracy": 0.6354625821113586,
+      "loss": 6.584984302520752,
+      "mean_token_accuracy": 0.6384725868701935,
       "num_tokens": 49872.0,
       "step": 7
     },
     {
-      "entropy": 1.2574806809425354,
+      "entropy": 1.175276130437851,
       "epoch": 0.20253164556962025,
-      "grad_norm": 17.659940719604492,
+      "grad_norm": 12.119710922241211,
       "learning_rate": 0.0001992320579737045,
-      "loss": 6.544322967529297,
-      "mean_token_accuracy": 0.626764640212059,
+      "loss": 6.310988426208496,
+      "mean_token_accuracy": 0.6375854164361954,
       "num_tokens": 55639.0,
       "step": 8
     },
     {
-      "entropy": 1.1509844064712524,
+      "entropy": 1.0149184465408325,
       "epoch": 0.22784810126582278,
-      "grad_norm": 8.395023345947266,
+      "grad_norm": 7.977464199066162,
       "learning_rate": 0.00019863613034027224,
-      "loss": 6.157719135284424,
-      "mean_token_accuracy": 0.669881597161293,
+      "loss": 6.036639213562012,
+      "mean_token_accuracy": 0.666093721985817,
       "num_tokens": 61649.0,
       "step": 9
     },
     {
-      "entropy": 1.0907686650753021,
+      "entropy": 1.022455170750618,
       "epoch": 0.25316455696202533,
-      "grad_norm": 4.606855392456055,
+      "grad_norm": 7.014733791351318,
       "learning_rate": 0.00019787168453273544,
-      "loss": 5.892803192138672,
-      "mean_token_accuracy": 0.6802202314138412,
+      "loss": 5.827225685119629,
+      "mean_token_accuracy": 0.6803832501173019,
       "num_tokens": 69689.0,
       "step": 10
     },
     {
-      "entropy": 1.1551010608673096,
+      "entropy": 1.1049264669418335,
       "epoch": 0.27848101265822783,
-      "grad_norm": 14.0298490524292,
+      "grad_norm": 5.545648574829102,
       "learning_rate": 0.00019694002659393305,
-      "loss": 5.6713948249816895,
-      "mean_token_accuracy": 0.6972689032554626,
+      "loss": 5.525129795074463,
+      "mean_token_accuracy": 0.704516813158989,
       "num_tokens": 77528.0,
       "step": 11
     },
     {
-      "entropy": 0.8588106781244278,
+      "entropy": 0.8492173552513123,
       "epoch": 0.3037974683544304,
-      "grad_norm": 4.56960916519165,
+      "grad_norm": 6.59067964553833,
       "learning_rate": 0.0001958427482458253,
-      "loss": 5.428313255310059,
-      "mean_token_accuracy": 0.7234913557767868,
+      "loss": 5.324199199676514,
+      "mean_token_accuracy": 0.8436015695333481,
       "num_tokens": 85064.0,
       "step": 12
     },
     {
-      "entropy": 0.705646425485611,
+      "entropy": 0.6838414371013641,
       "epoch": 0.3291139240506329,
-      "grad_norm": 3.344228744506836,
+      "grad_norm": 3.6837315559387207,
       "learning_rate": 0.00019458172417006347,
-      "loss": 5.444645404815674,
-      "mean_token_accuracy": 0.8130310028791428,
+      "loss": 5.420000076293945,
+      "mean_token_accuracy": 0.816602423787117,
       "num_tokens": 91009.0,
       "step": 13
     },
     {
-      "entropy": 0.7050664573907852,
+      "entropy": 0.6191327422857285,
       "epoch": 0.35443037974683544,
-      "grad_norm": 3.312098979949951,
+      "grad_norm": 3.5020408630371094,
       "learning_rate": 0.0001931591088051279,
-      "loss": 5.357623100280762,
-      "mean_token_accuracy": 0.8219081908464432,
+      "loss": 5.348326683044434,
+      "mean_token_accuracy": 0.8181768357753754,
       "num_tokens": 97909.0,
       "step": 14
     },
     {
-      "entropy": 0.6056851297616959,
+      "entropy": 0.5245884731411934,
       "epoch": 0.379746835443038,
-      "grad_norm": 4.054323196411133,
+      "grad_norm": 3.6154603958129883,
       "learning_rate": 0.00019157733266550575,
-      "loss": 5.329057216644287,
-      "mean_token_accuracy": 0.8183465003967285,
+      "loss": 5.308533191680908,
+      "mean_token_accuracy": 0.8219696879386902,
       "num_tokens": 104187.0,
       "step": 15
     },
     {
-      "entropy": 0.46965425461530685,
+      "entropy": 0.4121474325656891,
       "epoch": 0.4050632911392405,
-      "grad_norm": 3.1617722511291504,
+      "grad_norm": 2.822946071624756,
       "learning_rate": 0.0001898390981891979,
-      "loss": 5.159331321716309,
-      "mean_token_accuracy": 0.8249307423830032,
+      "loss": 5.135751247406006,
+      "mean_token_accuracy": 0.8321253657341003,
       "num_tokens": 110300.0,
       "step": 16
     },
     {
-      "entropy": 0.4227886497974396,
+      "entropy": 0.392690971493721,
       "epoch": 0.43037974683544306,
-      "grad_norm": 6.028562545776367,
+      "grad_norm": 5.22876501083374,
       "learning_rate": 0.0001879473751206489,
-      "loss": 5.351903915405273,
-      "mean_token_accuracy": 0.8241183459758759,
+      "loss": 5.34346866607666,
+      "mean_token_accuracy": 0.8164242208003998,
       "num_tokens": 117157.0,
       "step": 17
     },
     {
-      "entropy": 0.3527509272098541,
+      "entropy": 0.3356513977050781,
       "epoch": 0.45569620253164556,
-      "grad_norm": 3.293024778366089,
+      "grad_norm": 3.489384412765503,
       "learning_rate": 0.00018590539543698854,
-      "loss": 4.932947158813477,
-      "mean_token_accuracy": 0.8373657613992691,
+      "loss": 4.941544532775879,
+      "mean_token_accuracy": 0.8337943404912949,
       "num_tokens": 123378.0,
       "step": 18
     },
     {
-      "entropy": 0.33450157940387726,
+      "entropy": 0.32025614380836487,
       "epoch": 0.4810126582278481,
-      "grad_norm": 1.6703517436981201,
+      "grad_norm": 2.2713074684143066,
       "learning_rate": 0.00018371664782625287,
-      "loss": 5.134032249450684,
-      "mean_token_accuracy": 0.8269297778606415,
+      "loss": 5.1571879386901855,
+      "mean_token_accuracy": 0.8307176530361176,
       "num_tokens": 129042.0,
       "step": 19
     },
     {
-      "entropy": 0.3399146646261215,
+      "entropy": 0.3084551766514778,
       "epoch": 0.5063291139240507,
-      "grad_norm": 2.0219178199768066,
+      "grad_norm": 1.9176894426345825,
       "learning_rate": 0.0001813848717270195,
-      "loss": 5.027883052825928,
-      "mean_token_accuracy": 0.8322668075561523,
+      "loss": 5.014953136444092,
+      "mean_token_accuracy": 0.8466087132692337,
       "num_tokens": 135747.0,
       "step": 20
     },
     {
-      "entropy": 0.29102808609604836,
+      "entropy": 0.26904235407710075,
       "epoch": 0.5316455696202531,
-      "grad_norm": 1.350559949874878,
+      "grad_norm": 1.4043225049972534,
       "learning_rate": 0.00017891405093963938,
-      "loss": 4.954178810119629,
-      "mean_token_accuracy": 0.8439806550741196,
+      "loss": 4.949811935424805,
+      "mean_token_accuracy": 0.8474528640508652,
       "num_tokens": 141848.0,
       "step": 21
     },
     {
-      "entropy": 0.3136157840490341,
+      "entropy": 0.2922343239188194,
       "epoch": 0.5569620253164557,
-      "grad_norm": 1.3057941198349,
+      "grad_norm": 1.2163947820663452,
       "learning_rate": 0.00017630840681998066,
-      "loss": 5.179681301116943,
-      "mean_token_accuracy": 0.8438494354486465,
+      "loss": 5.169545650482178,
+      "mean_token_accuracy": 0.8475807756185532,
       "num_tokens": 147566.0,
       "step": 22
     },
     {
-      "entropy": 0.2950733080506325,
+      "entropy": 0.2782471589744091,
       "epoch": 0.5822784810126582,
-      "grad_norm": 1.1281788349151611,
+      "grad_norm": 1.0059943199157715,
       "learning_rate": 0.00017357239106731317,
-      "loss": 5.171978950500488,
+      "loss": 5.164263725280762,
       "mean_token_accuracy": 0.8462631553411484,
       "num_tokens": 154620.0,
       "step": 23
     },
     {
-      "entropy": 0.27676206082105637,
+      "entropy": 0.2586139217019081,
       "epoch": 0.6075949367088608,
-      "grad_norm": 1.311663269996643,
+      "grad_norm": 1.22340726852417,
       "learning_rate": 0.00017071067811865476,
-      "loss": 5.077099800109863,
+      "loss": 5.076096534729004,
       "mean_token_accuracy": 0.8611445128917694,
       "num_tokens": 161701.0,
       "step": 24
     },
     {
-      "entropy": 0.30840296298265457,
+      "entropy": 0.28559913113713264,
       "epoch": 0.6329113924050633,
-      "grad_norm": 2.281773090362549,
+      "grad_norm": 2.1334292888641357,
       "learning_rate": 0.00016772815716257412,
-      "loss": 5.118645668029785,
-      "mean_token_accuracy": 0.8323634117841721,
+      "loss": 5.121291160583496,
+      "mean_token_accuracy": 0.8396630734205246,
       "num_tokens": 168435.0,
       "step": 25
     },
     {
-      "entropy": 0.29603931307792664,
+      "entropy": 0.26606928184628487,
       "epoch": 0.6582278481012658,
-      "grad_norm": 1.230368971824646,
+      "grad_norm": 0.8918235301971436,
       "learning_rate": 0.00016462992378609407,
-      "loss": 5.176095962524414,
-      "mean_token_accuracy": 0.839859664440155,
+      "loss": 5.148040771484375,
+      "mean_token_accuracy": 0.8437058180570602,
       "num_tokens": 175650.0,
       "step": 26
     },
     {
-      "entropy": 0.24447191506624222,
+      "entropy": 0.23234782740473747,
       "epoch": 0.6835443037974683,
-      "grad_norm": 1.7844265699386597,
+      "grad_norm": 1.8136335611343384,
       "learning_rate": 0.0001614212712689668,
-      "loss": 5.146295547485352,
-      "mean_token_accuracy": 0.8507861793041229,
+      "loss": 5.150457382202148,
+      "mean_token_accuracy": 0.8433783650398254,
       "num_tokens": 181490.0,
       "step": 27
     },
     {
-      "entropy": 0.23978285491466522,
+      "entropy": 0.24661532044410706,
       "epoch": 0.7088607594936709,
-      "grad_norm": 1.4895042181015015,
+      "grad_norm": 1.321523904800415,
       "learning_rate": 0.00015810768154019385,
-      "loss": 5.0867919921875,
-      "mean_token_accuracy": 0.8498348891735077,
+      "loss": 5.084752082824707,
+      "mean_token_accuracy": 0.8424803614616394,
       "num_tokens": 188582.0,
       "step": 28
     },
     {
-      "entropy": 0.1993725746870041,
+      "entropy": 0.19287967681884766,
       "epoch": 0.7341772151898734,
-      "grad_norm": 1.2189953327178955,
+      "grad_norm": 0.9798034429550171,
       "learning_rate": 0.00015469481581224272,
-      "loss": 5.131974697113037,
-      "mean_token_accuracy": 0.8698419630527496,
+      "loss": 5.147141456604004,
+      "mean_token_accuracy": 0.866110622882843,
       "num_tokens": 193783.0,
       "step": 29
     },
     {
-      "entropy": 0.24254225194454193,
+      "entropy": 0.23315827921032906,
       "epoch": 0.759493670886076,
-      "grad_norm": 3.0398507118225098,
+      "grad_norm": 3.1756558418273926,
       "learning_rate": 0.00015118850490896012,
-      "loss": 5.307463645935059,
-      "mean_token_accuracy": 0.8126509040594101,
+      "loss": 5.3330183029174805,
+      "mean_token_accuracy": 0.8204634040594101,
       "num_tokens": 200686.0,
       "step": 30
     },
     {
-      "entropy": 0.18977511674165726,
+      "entropy": 0.1938694752752781,
       "epoch": 0.7848101265822784,
-      "grad_norm": 1.7506808042526245,
+      "grad_norm": 1.9134807586669922,
       "learning_rate": 0.00014759473930370736,
-      "loss": 5.047242164611816,
-      "mean_token_accuracy": 0.8546684086322784,
+      "loss": 5.052967071533203,
+      "mean_token_accuracy": 0.8582398444414139,
       "num_tokens": 208440.0,
       "step": 31
     },
     {
-      "entropy": 0.2233816385269165,
+      "entropy": 0.22125443816184998,
       "epoch": 0.810126582278481,
-      "grad_norm": 1.0086487531661987,
+      "grad_norm": 1.5088441371917725,
       "learning_rate": 0.00014391965888473703,
-      "loss": 5.027800559997559,
-      "mean_token_accuracy": 0.8576832860708237,
+      "loss": 5.049187660217285,
+      "mean_token_accuracy": 0.8358375877141953,
       "num_tokens": 215131.0,
       "step": 32
     },
     {
-      "entropy": 0.19872410595417023,
+      "entropy": 0.20639389753341675,
       "epoch": 0.8354430379746836,
-      "grad_norm": 1.6391977071762085,
+      "grad_norm": 1.491873860359192,
       "learning_rate": 0.00014016954246529696,
-      "loss": 5.0654706954956055,
-      "mean_token_accuracy": 0.8516139537096024,
+      "loss": 5.055438041687012,
+      "mean_token_accuracy": 0.8552371561527252,
       "num_tokens": 222195.0,
       "step": 33
     },
     {
-      "entropy": 0.20351751521229744,
+      "entropy": 0.20158018916845322,
       "epoch": 0.8607594936708861,
-      "grad_norm": 1.5714291334152222,
+      "grad_norm": 1.46501624584198,
       "learning_rate": 0.00013635079705638298,
-      "loss": 5.170158386230469,
-      "mean_token_accuracy": 0.8566406965255737,
+      "loss": 5.16477108001709,
+      "mean_token_accuracy": 0.8602638840675354,
       "num_tokens": 229926.0,
       "step": 34
     },
     {
-      "entropy": 0.21749146282672882,
+      "entropy": 0.21367954835295677,
       "epoch": 0.8860759493670886,
-      "grad_norm": 1.6517112255096436,
+      "grad_norm": 1.7227668762207031,
       "learning_rate": 0.00013246994692046836,
-      "loss": 5.215228080749512,
+      "loss": 5.240202903747559,
       "mean_token_accuracy": 0.8288360238075256,
       "num_tokens": 235766.0,
       "step": 35
     },
     {
-      "entropy": 0.2011272795498371,
+      "entropy": 0.2060464285314083,
       "epoch": 0.9113924050632911,
-      "grad_norm": 1.4017918109893799,
+      "grad_norm": 1.615714192390442,
       "learning_rate": 0.00012853362242491053,
-      "loss": 5.1320648193359375,
+      "loss": 5.13973331451416,
       "mean_token_accuracy": 0.8528883457183838,
       "num_tokens": 242731.0,
       "step": 36
     },
     {
-      "entropy": 0.21461713686585426,
+      "entropy": 0.2112499214708805,
       "epoch": 0.9367088607594937,
-      "grad_norm": 1.77144193649292,
+      "grad_norm": 1.70606529712677,
       "learning_rate": 0.00012454854871407994,
-      "loss": 5.11080265045166,
-      "mean_token_accuracy": 0.8521855026483536,
+      "loss": 5.105003833770752,
+      "mean_token_accuracy": 0.8559168428182602,
       "num_tokens": 249818.0,
       "step": 37
     },
     {
-      "entropy": 0.19789887592196465,
+      "entropy": 0.1963396780192852,
       "epoch": 0.9620253164556962,
-      "grad_norm": 3.2191762924194336,
+      "grad_norm": 1.8689205646514893,
       "learning_rate": 0.00012052153421956342,
-      "loss": 5.070086479187012,
-      "mean_token_accuracy": 0.848893016576767,
+      "loss": 5.083700180053711,
+      "mean_token_accuracy": 0.8450468629598618,
       "num_tokens": 256043.0,
       "step": 38
     },
     {
-      "entropy": 0.1783522628247738,
+      "entropy": 0.17459847778081894,
       "epoch": 0.9873417721518988,
-      "grad_norm": 1.3007391691207886,
+      "grad_norm": 1.3199875354766846,
       "learning_rate": 0.00011645945902807341,
-      "loss": 5.1378865242004395,
-      "mean_token_accuracy": 0.8583788871765137,
+      "loss": 5.136896133422852,
+      "mean_token_accuracy": 0.8544726371765137,
       "num_tokens": 263628.0,
       "step": 39
     },
     {
-      "entropy": 0.21517179906368256,
+      "entropy": 0.19893981516361237,
       "epoch": 1.0,
-      "grad_norm": 2.357231378555298,
+      "grad_norm": 2.03135347366333,
       "learning_rate": 0.00011236926312693479,
-      "loss": 5.041143894195557,
-      "mean_token_accuracy": 0.8405462205410004,
+      "loss": 5.0191426277160645,
+      "mean_token_accuracy": 0.8476890921592712,
       "num_tokens": 266935.0,
       "step": 40
     },
@@ -415,407 +415,407 @@
       "eval_loss": NaN,
       "eval_mean_token_accuracy": 0.0,
       "eval_num_tokens": 266935.0,
-      "eval_runtime": 5.9781,
-      "eval_samples_per_second": 37.136,
-      "eval_steps_per_second": 9.368,
+      "eval_runtime": 5.2068,
+      "eval_samples_per_second": 42.637,
+      "eval_steps_per_second": 10.755,
       "step": 40
     },
     {
-      "entropy": 0.1771511249244213,
+      "entropy": 0.17801347747445107,
       "epoch": 1.0253164556962024,
-      "grad_norm": 1.8927557468414307,
+      "grad_norm": 2.8673174381256104,
       "learning_rate": 0.00010825793454723325,
-      "loss": 5.032665252685547,
-      "mean_token_accuracy": 0.8413392454385757,
+      "loss": 5.022855281829834,
+      "mean_token_accuracy": 0.8449043035507202,
       "num_tokens": 272689.0,
       "step": 41
     },
     {
-      "entropy": 0.19083253294229507,
+      "entropy": 0.18583980947732925,
       "epoch": 1.0506329113924051,
-      "grad_norm": 1.5800896883010864,
+      "grad_norm": 1.413786768913269,
       "learning_rate": 0.00010413249742488131,
-      "loss": 5.118054389953613,
+      "loss": 5.111584663391113,
       "mean_token_accuracy": 0.8499636799097061,
       "num_tokens": 279320.0,
       "step": 42
     },
     {
-      "entropy": 0.1893175132572651,
+      "entropy": 0.18669037148356438,
       "epoch": 1.0759493670886076,
-      "grad_norm": 1.6944535970687866,
+      "grad_norm": 1.6076576709747314,
       "learning_rate": 0.0001,
-      "loss": 4.96186637878418,
-      "mean_token_accuracy": 0.846690446138382,
+      "loss": 4.956506729125977,
+      "mean_token_accuracy": 0.8430139720439911,
       "num_tokens": 285858.0,
       "step": 43
     },
     {
-      "entropy": 0.19514397159218788,
+      "entropy": 0.1918954849243164,
       "epoch": 1.1012658227848102,
-      "grad_norm": 1.0285674333572388,
+      "grad_norm": 0.9509457349777222,
       "learning_rate": 9.586750257511867e-05,
-      "loss": 5.133388519287109,
+      "loss": 5.113883972167969,
       "mean_token_accuracy": 0.8598480075597763,
       "num_tokens": 293542.0,
       "step": 44
     },
     {
-      "entropy": 0.1813015267252922,
+      "entropy": 0.18484549596905708,
       "epoch": 1.1265822784810127,
-      "grad_norm": 0.7471196055412292,
+      "grad_norm": 0.8191195726394653,
       "learning_rate": 9.174206545276677e-05,
-      "loss": 5.058955192565918,
+      "loss": 5.065518379211426,
       "mean_token_accuracy": 0.8599645048379898,
       "num_tokens": 300567.0,
       "step": 45
     },
     {
-      "entropy": 0.19578944519162178,
+      "entropy": 0.20070786401629448,
       "epoch": 1.1518987341772151,
-      "grad_norm": 1.092103123664856,
+      "grad_norm": 1.3238853216171265,
       "learning_rate": 8.763073687306524e-05,
-      "loss": 5.134239196777344,
-      "mean_token_accuracy": 0.8482615053653717,
+      "loss": 5.136753082275391,
+      "mean_token_accuracy": 0.8521076589822769,
       "num_tokens": 308202.0,
       "step": 46
     },
     {
-      "entropy": 0.20517433807253838,
+      "entropy": 0.20315109938383102,
       "epoch": 1.1772151898734178,
-      "grad_norm": 1.2428195476531982,
+      "grad_norm": 1.067439079284668,
       "learning_rate": 8.35405409719266e-05,
-      "loss": 5.1778974533081055,
-      "mean_token_accuracy": 0.8619973361492157,
+      "loss": 5.177618026733398,
+      "mean_token_accuracy": 0.8582659959793091,
       "num_tokens": 314689.0,
       "step": 47
     },
     {
-      "entropy": 0.19207627698779106,
+      "entropy": 0.1983315795660019,
       "epoch": 1.2025316455696202,
-      "grad_norm": 1.1919926404953003,
+      "grad_norm": 1.124258279800415,
       "learning_rate": 7.947846578043659e-05,
-      "loss": 5.037685394287109,
+      "loss": 5.045456409454346,
       "mean_token_accuracy": 0.8561878055334091,
       "num_tokens": 320323.0,
       "step": 48
     },
     {
-      "entropy": 0.1860572211444378,
+      "entropy": 0.1914753057062626,
       "epoch": 1.2278481012658227,
-      "grad_norm": 2.1096537113189697,
+      "grad_norm": 2.017146110534668,
       "learning_rate": 7.54514512859201e-05,
-      "loss": 5.031648635864258,
-      "mean_token_accuracy": 0.8480775505304337,
+      "loss": 5.0167694091796875,
+      "mean_token_accuracy": 0.8443461954593658,
       "num_tokens": 327016.0,
       "step": 49
     },
     {
-      "entropy": 0.1999019868671894,
+      "entropy": 0.19617706537246704,
       "epoch": 1.2531645569620253,
-      "grad_norm": 1.2548681497573853,
+      "grad_norm": 1.0709376335144043,
       "learning_rate": 7.146637757508949e-05,
-      "loss": 5.141640663146973,
-      "mean_token_accuracy": 0.865373820066452,
+      "loss": 5.132477760314941,
+      "mean_token_accuracy": 0.8692199736833572,
       "num_tokens": 333272.0,
       "step": 50
     },
     {
-      "entropy": 0.20417018979787827,
+      "entropy": 0.2076304629445076,
       "epoch": 1.2784810126582278,
-      "grad_norm": 0.847514808177948,
+      "grad_norm": 0.8743495345115662,
       "learning_rate": 6.753005307953167e-05,
-      "loss": 5.055359840393066,
+      "loss": 5.0548014640808105,
       "mean_token_accuracy": 0.8717339187860489,
       "num_tokens": 340919.0,
       "step": 51
     },
     {
-      "entropy": 0.20846642926335335,
+      "entropy": 0.20958010107278824,
       "epoch": 1.3037974683544304,
-      "grad_norm": 1.305163025856018,
+      "grad_norm": 1.33949875831604,
       "learning_rate": 6.3649202943617e-05,
-      "loss": 5.115406513214111,
-      "mean_token_accuracy": 0.8486459404230118,
+      "loss": 5.117767333984375,
+      "mean_token_accuracy": 0.8488624095916748,
       "num_tokens": 347473.0,
       "step": 52
     },
     {
-      "entropy": 0.2167152315378189,
+      "entropy": 0.21564103290438652,
       "epoch": 1.3291139240506329,
-      "grad_norm": 1.275968074798584,
+      "grad_norm": 1.3259762525558472,
       "learning_rate": 5.983045753470308e-05,
-      "loss": 5.049930095672607,
-      "mean_token_accuracy": 0.8500020503997803,
+      "loss": 5.061050891876221,
+      "mean_token_accuracy": 0.8425844609737396,
       "num_tokens": 353979.0,
       "step": 53
     },
     {
-      "entropy": 0.1877211444079876,
+      "entropy": 0.18632080778479576,
       "epoch": 1.3544303797468356,
-      "grad_norm": 1.289473533630371,
+      "grad_norm": 1.2237317562103271,
       "learning_rate": 5.608034111526298e-05,
-      "loss": 5.212080001831055,
+      "loss": 5.200876235961914,
       "mean_token_accuracy": 0.8610779196023941,
       "num_tokens": 360939.0,
       "step": 54
     },
     {
-      "entropy": 0.21858176216483116,
+      "entropy": 0.21953574940562248,
       "epoch": 1.379746835443038,
-      "grad_norm": 1.2420766353607178,
+      "grad_norm": 1.2576055526733398,
       "learning_rate": 5.240526069629265e-05,
-      "loss": 5.145570278167725,
+      "loss": 5.142441749572754,
       "mean_token_accuracy": 0.8502416759729385,
       "num_tokens": 367684.0,
       "step": 55
     },
     {
-      "entropy": 0.16584616154432297,
+      "entropy": 0.16833588108420372,
       "epoch": 1.4050632911392404,
-      "grad_norm": 0.8850894570350647,
+      "grad_norm": 0.8653225302696228,
       "learning_rate": 4.8811495091039926e-05,
-      "loss": 4.982363224029541,
+      "loss": 4.981797218322754,
       "mean_token_accuracy": 0.8663530200719833,
       "num_tokens": 374983.0,
       "step": 56
     },
     {
-      "entropy": 0.20285014435648918,
+      "entropy": 0.19686222821474075,
       "epoch": 1.4303797468354431,
-      "grad_norm": 0.9019045233726501,
+      "grad_norm": 0.9341261386871338,
       "learning_rate": 4.530518418775733e-05,
-      "loss": 5.186882972717285,
-      "mean_token_accuracy": 0.8562371581792831,
+      "loss": 5.197404861450195,
+      "mean_token_accuracy": 0.8523910045623779,
       "num_tokens": 380947.0,
       "step": 57
     },
     {
-      "entropy": 0.19403913989663124,
+      "entropy": 0.18849175423383713,
       "epoch": 1.4556962025316456,
-      "grad_norm": 1.3603911399841309,
+      "grad_norm": 1.3819215297698975,
       "learning_rate": 4.189231845980618e-05,
-      "loss": 5.046579837799072,
+      "loss": 5.049962997436523,
       "mean_token_accuracy": 0.856428250670433,
       "num_tokens": 388400.0,
       "step": 58
     },
     {
-      "entropy": 0.17723571881651878,
+      "entropy": 0.17277341708540916,
       "epoch": 1.481012658227848,
-      "grad_norm": 0.8954723477363586,
+      "grad_norm": 0.7869784832000732,
       "learning_rate": 3.857872873103322e-05,
-      "loss": 4.9259443283081055,
-      "mean_token_accuracy": 0.8559762090444565,
+      "loss": 4.933191299438477,
+      "mean_token_accuracy": 0.8741099387407303,
       "num_tokens": 394297.0,
       "step": 59
     },
     {
-      "entropy": 0.1926303431391716,
+      "entropy": 0.191244438290596,
       "epoch": 1.5063291139240507,
-      "grad_norm": 1.165939211845398,
+      "grad_norm": 1.2609248161315918,
       "learning_rate": 3.53700762139059e-05,
-      "loss": 5.209965705871582,
-      "mean_token_accuracy": 0.8545413017272949,
+      "loss": 5.227235794067383,
+      "mean_token_accuracy": 0.8469072580337524,
       "num_tokens": 401531.0,
       "step": 60
     },
     {
-      "entropy": 0.17567158862948418,
+      "entropy": 0.17449039965867996,
       "epoch": 1.5316455696202531,
-      "grad_norm": 1.256475806236267,
+      "grad_norm": 1.4421076774597168,
       "learning_rate": 3.227184283742591e-05,
-      "loss": 5.071780681610107,
+      "loss": 5.084752559661865,
       "mean_token_accuracy": 0.8629807680845261,
       "num_tokens": 408001.0,
       "step": 61
     },
     {
-      "entropy": 0.15778229013085365,
+      "entropy": 0.15809274092316628,
       "epoch": 1.5569620253164556,
-      "grad_norm": 0.9394585490226746,
+      "grad_norm": 1.0014944076538086,
       "learning_rate": 2.9289321881345254e-05,
-      "loss": 5.174696445465088,
+      "loss": 5.163921356201172,
       "mean_token_accuracy": 0.8652869164943695,
       "num_tokens": 414848.0,
       "step": 62
     },
     {
-      "entropy": 0.2071729153394699,
+      "entropy": 0.19591599330306053,
       "epoch": 1.5822784810126582,
-      "grad_norm": 0.9868037700653076,
+      "grad_norm": 0.9324407577514648,
       "learning_rate": 2.6427608932686843e-05,
-      "loss": 5.067627906799316,
-      "mean_token_accuracy": 0.8562931418418884,
+      "loss": 5.06367301940918,
+      "mean_token_accuracy": 0.8634967058897018,
       "num_tokens": 422561.0,
       "step": 63
     },
     {
-      "entropy": 0.1966983526945114,
+      "entropy": 0.19570430740714073,
       "epoch": 1.6075949367088609,
-      "grad_norm": 0.7985685467720032,
+      "grad_norm": 0.7240243554115295,
       "learning_rate": 2.3691593180019366e-05,
-      "loss": 5.114544868469238,
-      "mean_token_accuracy": 0.8560546040534973,
+      "loss": 5.119420051574707,
+      "mean_token_accuracy": 0.859731063246727,
       "num_tokens": 428474.0,
       "step": 64
     },
     {
-      "entropy": 0.1941680982708931,
+      "entropy": 0.18400323763489723,
       "epoch": 1.6329113924050633,
-      "grad_norm": 1.1260875463485718,
+      "grad_norm": 1.211584448814392,
       "learning_rate": 2.1085949060360654e-05,
-      "loss": 5.048023223876953,
-      "mean_token_accuracy": 0.8535272777080536,
+      "loss": 5.057708263397217,
+      "mean_token_accuracy": 0.8534739911556244,
       "num_tokens": 433867.0,
       "step": 65
     },
     {
-      "entropy": 0.18942571431398392,
+      "entropy": 0.1828528456389904,
       "epoch": 1.6582278481012658,
-      "grad_norm": 0.8621047139167786,
+      "grad_norm": 0.8080496191978455,
       "learning_rate": 1.861512827298051e-05,
-      "loss": 4.9987688064575195,
-      "mean_token_accuracy": 0.8621815890073776,
+      "loss": 4.988627910614014,
+      "mean_token_accuracy": 0.8694416880607605,
       "num_tokens": 440685.0,
       "step": 66
     },
     {
-      "entropy": 0.18620551750063896,
+      "entropy": 0.18491194397211075,
       "epoch": 1.6835443037974684,
-      "grad_norm": 0.8113185167312622,
+      "grad_norm": 0.8978958129882812,
       "learning_rate": 1.6283352173747145e-05,
-      "loss": 5.130119323730469,
+      "loss": 5.144866943359375,
       "mean_token_accuracy": 0.8656566739082336,
       "num_tokens": 447730.0,
       "step": 67
     },
     {
-      "entropy": 0.1812417395412922,
+      "entropy": 0.18067174404859543,
       "epoch": 1.7088607594936709,
-      "grad_norm": 0.657331645488739,
+      "grad_norm": 0.6968832015991211,
       "learning_rate": 1.4094604563011472e-05,
-      "loss": 5.11595344543457,
+      "loss": 5.13330078125,
       "mean_token_accuracy": 0.8689732104539871,
       "num_tokens": 454057.0,
       "step": 68
     },
     {
-      "entropy": 0.17315322160720825,
+      "entropy": 0.16753962635993958,
       "epoch": 1.7341772151898733,
-      "grad_norm": 1.2474181652069092,
+      "grad_norm": 1.267209768295288,
       "learning_rate": 1.2052624879351104e-05,
-      "loss": 5.142193794250488,
+      "loss": 5.151784896850586,
       "mean_token_accuracy": 0.8530041426420212,
       "num_tokens": 460989.0,
       "step": 69
     },
     {
-      "entropy": 0.20847681164741516,
+      "entropy": 0.20549893751740456,
       "epoch": 1.759493670886076,
-      "grad_norm": 0.9265520572662354,
+      "grad_norm": 0.9581330418586731,
       "learning_rate": 1.0160901810802115e-05,
-      "loss": 4.949027061462402,
-      "mean_token_accuracy": 0.8633775115013123,
+      "loss": 4.949734687805176,
+      "mean_token_accuracy": 0.8597010374069214,
       "num_tokens": 468216.0,
       "step": 70
     },
     {
-      "entropy": 0.1857427917420864,
+      "entropy": 0.18367141112685204,
       "epoch": 1.7848101265822784,
-      "grad_norm": 0.8429440259933472,
+      "grad_norm": 0.9142661690711975,
       "learning_rate": 8.422667334494249e-06,
-      "loss": 5.067360877990723,
-      "mean_token_accuracy": 0.8449466675519943,
+      "loss": 5.066184997558594,
+      "mean_token_accuracy": 0.8487345576286316,
       "num_tokens": 474276.0,
       "step": 71
     },
     {
-      "entropy": 0.17809578776359558,
+      "entropy": 0.17542577907443047,
       "epoch": 1.810126582278481,
-      "grad_norm": 1.0042821168899536,
+      "grad_norm": 0.9961230158805847,
       "learning_rate": 6.840891194872112e-06,
-      "loss": 5.115839958190918,
+      "loss": 5.105713844299316,
       "mean_token_accuracy": 0.8699368089437485,
       "num_tokens": 481319.0,
       "step": 72
     },
     {
-      "entropy": 0.18285344913601875,
+      "entropy": 0.17863155156373978,
       "epoch": 1.8354430379746836,
-      "grad_norm": 0.9699199199676514,
+      "grad_norm": 1.0096696615219116,
       "learning_rate": 5.418275829936537e-06,
-      "loss": 5.113241672515869,
-      "mean_token_accuracy": 0.8597128540277481,
+      "loss": 5.117948055267334,
+      "mean_token_accuracy": 0.8560896664857864,
       "num_tokens": 488560.0,
       "step": 73
     },
     {
-      "entropy": 0.19144808873534203,
+      "entropy": 0.19148987531661987,
       "epoch": 1.8607594936708862,
-      "grad_norm": 0.7952200174331665,
+      "grad_norm": 0.8309740424156189,
       "learning_rate": 4.1572517541747294e-06,
-      "loss": 5.170323371887207,
+      "loss": 5.165389060974121,
       "mean_token_accuracy": 0.86890609562397,
       "num_tokens": 494057.0,
       "step": 74
     },
     {
-      "entropy": 0.18293721601366997,
+      "entropy": 0.180438794195652,
       "epoch": 1.8860759493670884,
-      "grad_norm": 0.8813140392303467,
+      "grad_norm": 0.8403601050376892,
       "learning_rate": 3.059973406066963e-06,
-      "loss": 4.910680770874023,
-      "mean_token_accuracy": 0.867401659488678,
+      "loss": 4.906594276428223,
+      "mean_token_accuracy": 0.8745459914207458,
       "num_tokens": 501562.0,
       "step": 75
     },
     {
-      "entropy": 0.1984117217361927,
+      "entropy": 0.19829363003373146,
       "epoch": 1.9113924050632911,
-      "grad_norm": 1.1274702548980713,
+      "grad_norm": 1.2034733295440674,
       "learning_rate": 2.128315467264552e-06,
-      "loss": 5.088738441467285,
-      "mean_token_accuracy": 0.8530780524015427,
+      "loss": 5.0864434242248535,
+      "mean_token_accuracy": 0.849346712231636,
       "num_tokens": 509439.0,
       "step": 76
     },
     {
-      "entropy": 0.18629329279065132,
+      "entropy": 0.18359657377004623,
       "epoch": 1.9367088607594938,
-      "grad_norm": 0.9891031980514526,
+      "grad_norm": 1.079089879989624,
       "learning_rate": 1.3638696597277679e-06,
-      "loss": 5.085946083068848,
+      "loss": 5.085855484008789,
       "mean_token_accuracy": 0.8562160283327103,
       "num_tokens": 516267.0,
       "step": 77
     },
     {
-      "entropy": 0.16653455793857574,
+      "entropy": 0.16170304268598557,
       "epoch": 1.9620253164556962,
-      "grad_norm": 0.746681809425354,
+      "grad_norm": 0.7529839873313904,
       "learning_rate": 7.679420262954984e-07,
-      "loss": 5.034550189971924,
+      "loss": 5.025687217712402,
       "mean_token_accuracy": 0.8639013916254044,
       "num_tokens": 523626.0,
       "step": 78
     },
     {
-      "entropy": 0.1866554580628872,
+      "entropy": 0.18417693302035332,
       "epoch": 1.9873417721518987,
-      "grad_norm": 0.9339231848716736,
+      "grad_norm": 0.9824067950248718,
       "learning_rate": 3.415506993330153e-07,
-      "loss": 5.026074409484863,
+      "loss": 5.028251647949219,
       "mean_token_accuracy": 0.8578499257564545,
       "num_tokens": 530022.0,
       "step": 79
     },
     {
-      "entropy": 0.19535863399505615,
+      "entropy": 0.1983117312192917,
       "epoch": 2.0,
-      "grad_norm": 1.5845301151275635,
+      "grad_norm": 1.6459767818450928,
       "learning_rate": 8.542416126989805e-08,
-      "loss": 5.070830821990967,
+      "loss": 5.08328914642334,
       "mean_token_accuracy": 0.8686061203479767,
       "num_tokens": 533275.0,
       "step": 80
@@ -826,9 +826,9 @@
       "eval_loss": NaN,
       "eval_mean_token_accuracy": 0.0,
       "eval_num_tokens": 533275.0,
-      "eval_runtime": 5.983,
-      "eval_samples_per_second": 37.105,
-      "eval_steps_per_second": 9.36,
+      "eval_runtime": 5.2672,
+      "eval_samples_per_second": 42.148,
+      "eval_steps_per_second": 10.632,
       "step": 80
     }
   ],
@@ -858,7 +858,7 @@
       "attributes": {}
     }
   },
-  "total_flos": 1119703283007488.0,
+  "total_flos": 1115809521860608.0,
   "train_batch_size": 2,
   "trial_name": null,
   "trial_params": null

checkpoint-80/training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:da3e7916a9c3a9bdc1d65a3a86e68c6c06ab04fd280e02675e7357542392c8fa
+oid sha256:b5aae0a2160af9bfafbf734dc2e956a27d2ec4ded9101f3494b0466f6c3c55ff
 size 5496

predictions.jsonl

@@ -11,21 +11,21 @@
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n### Summary Missing validation logic in the storage volume import logic allows an authenticated user with access to Incus' storage volume feature to cause the Incus daemon to crash. Repeated use of this issue can be used to keep Incus offline causing a denial of service. ### Details The backup restore subsystem contains an out-of-bounds panic vulnerability caused by an invalid bounds check when indexing snapshot metadata arrays. The same flawed pattern also appears in the migration path. When iterating through physical snapshots provided in a backup archive, the loop uses the index i to look up corresponding metadata in the parsed Config.Snapshots and Config.VolumeSnapshots slices. To ensure that the metadata slice is long enough, the code uses the guard condition len(slice) >= i-1. This check is incorrect because it can still evaluate to true when the subsequent slice[i] access is out of bounds, including when i >= len(slice), triggering a runtime panic. An attacker can trigger this by submitting a backup archive that contains physical snapshot directories, which drive the loop variable i, while supplying a tampered index.yaml with an empty or truncated snapshot metadata array. This causes the daemon to index beyond the end of the metadata slice and crash, resulting in immediate denial of service on the node. Affected File: https://github.com/lxc/incus/blob/v6.22.0/internal/server/storage/backend.go Affected Code: ``` func (b *backend) CreateInstanceFromBackup(srcBacku\n</incident>"}], "ground_truth": "dos", "generated_content": "dos"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nAIAAIC report: Republicans support Trump using AI-generated kitten and duck images. Technology: Generative AI. Purpose: Defend reputation. Ethical issues: Mis/disinformation; Transparency.\n</incident>"}], "ground_truth": "misinformation", "generated_content": "misinformation"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n### Summary An authentication bypass vulnerability in phpMyFAQ allows any unauthenticated attacker to reset the password of any user account, including SuperAdmin accounts. By sending a PUT request with just a valid username and associated email address to /api/user/password/update, an attacker receives a new plaintext password via email without any token verification, rate limiting, or email confirmation. This enables complete account takeover of any user, including full administrative access. ### Details File: phpmyfaq/src/phpMyFAQ/Controller/Frontend/Api/UnauthorizedUserController.php Lines: 56-130 The updatePassword() method at line 56 accepts PUT requests to /user/password/update with only username and email in the JSON body: #[Route(path: 'user/password/update', name: 'api.private.user.password', methods: ['PUT'])] ```php public function updatePassword(Request $request): JsonResponse { $data = json_decode($request->getContent()); $username = trim((string) Filter::filterVar($data->username, FILTER_SANITIZE_SPECIAL_CHARS)); $email = trim((string) Filter::filterEmail($data->email)); if ($username !== '' && $username !== '0' && ($email !== '' && $email !== '0')) { $user = ($this->currentUserFactory ?? CurrentUser::getCurrentUser(...))($this->configuration); $loginExist = $user->getUserByLogin($username); if ($loginExist && $email === $user->getUserData('email')) { // NO TOKEN CHECK // NO RATE LIMITING\n</incident>"}], "ground_truth": "auth-bypass", "generated_content": "auth-bypass"}
-{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nn8n-MCP is an MCP server that provides AI assistants access to n8n node documentation, properties, and operations. Prior to version 2.47.11, when n8n-mcp runs in HTTP transport mode, incoming requests to the POST /mcp endpoint had their request metadata written to server logs regardless of the authentication outcome. In deployments where logs are collected, forwarded to external systems, or viewable outside the request trust boundary (shared log storage, SIEM pipelines, support/ops access), this can result in disclosure of: bearer tokens from the Authorization header, per-tenant API keys from the, x-n8n-key header in multi-tenant setups, JSON-RPC request payloads sent to the MCP endpoint. Access control itself was not bypassed \u2014 unauthenticated requests were correctly rejected with 401 Unauthorized \u2014 but sensitive values from those rejected requests could still be persisted in logs. This issue has been patched in version 2.47.11.\n</incident>"}], "ground_truth": "auth-bypass", "generated_content": "auth-bypass"}
+{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nn8n-MCP is an MCP server that provides AI assistants access to n8n node documentation, properties, and operations. Prior to version 2.47.11, when n8n-mcp runs in HTTP transport mode, incoming requests to the POST /mcp endpoint had their request metadata written to server logs regardless of the authentication outcome. In deployments where logs are collected, forwarded to external systems, or viewable outside the request trust boundary (shared log storage, SIEM pipelines, support/ops access), this can result in disclosure of: bearer tokens from the Authorization header, per-tenant API keys from the, x-n8n-key header in multi-tenant setups, JSON-RPC request payloads sent to the MCP endpoint. Access control itself was not bypassed \u2014 unauthenticated requests were correctly rejected with 401 Unauthorized \u2014 but sensitive values from those rejected requests could still be persisted in logs. This issue has been patched in version 2.47.11.\n</incident>"}], "ground_truth": "auth-bypass", "generated_content": "info-disclosure"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n### Summary diesel-async exposes uninitialized stack padding to safe code on every read of a MySQL `DATE`, `TIME`, `DATETIME`, or `TIMESTAMP` column. Reading that buffer is undefined behavior, and the leaked bytes can contain stale heap/stack contents, so this is both a soundness bug and a potential information-disclosure vector. ### Details In `diesel-async/src/mysql/row.rs` (lines 65-103), `MysqlRow::get` builds a `MysqlTime` from the parsed `mysql_async::Value` and then fabricates the byte buffer that downstream `FromSql` impls expect like this: ```rust let date = MysqlTime::new(/* fields from Value::Date / Value::Time */); let buffer = unsafe { let ptr = &date as *const MysqlTime as *const u8; let slice = std::slice::from_raw_parts(ptr, std::mem::size_of::<MysqlTime>()); slice.to_vec() }; ``` `MysqlTime` is `#[repr(C)]` with 3 bytes of padding after `bool neg` (Linux x86_64, offsets 0x21..0x23). The literal construction leaves that padding uninitialized, and `to_vec()` carries it into a `Vec<u8>` that becomes the `MysqlValue`'s backing buffer, reachable from safe code via `MysqlValue::as_bytes() -> &[u8]`. `diesel` itself avoids this by going through `MaybeUninit::<MysqlTime>::zeroed()` + `ptr::copy_nonoverlapping` (see `diesel/src/mysql/value.rs:43-94`); the same pattern would fix this. Alternatively, write the bytes diesel's `FromSql` reads without round-tripping through a `MysqlTime` value. ### PoC `Cargo.toml`: ```toml [dependencies] diesel = { ver\n</incident>"}], "ground_truth": "info-disclosure", "generated_content": "info-disclosure"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n## Summary The `list_files()` tool in `FileTools` validates the `directory` parameter against workspace boundaries via `_validate_path()`, but passes the `pattern` parameter directly to `Path.glob()` without any validation. Since Python's `Path.glob()` supports `..` path segments, an attacker can use relative path traversal in the glob pattern to enumerate arbitrary files outside the workspace, obtaining file metadata (existence, name, size, timestamps) for any path on the filesystem. ## Details The `_validate_path()` method at `file_tools.py:25` correctly prevents path traversal by checking for `..` segments and verifying the resolved path falls within the current workspace. All file operations (`read_file`, `write_file`, `copy_file`, etc.) route through this validation. However, `list_files()` at `file_tools.py:114` only validates the `directory` parameter (line 127), while the `pattern` parameter is passed directly to `Path.glob()` on line 130: ```python @staticmethod def list_files(directory: str, pattern: Optional[str] = None) -> List[Dict[str, Union[str, int]]]: try: safe_dir = FileTools._validate_path(directory) # directory validated path = Path(safe_dir) if pattern: files = path.glob(pattern) # pattern NOT validated \u2014 traversal possible else: files = path.iterdir() result = [] for file in files: if file.is_file(): stat = file.stat() result.ap\n</incident>"}], "ground_truth": "prompt-injection", "generated_content": "path-traversal"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nMicrosoft allowed contractors in China to access and review thousands of Skype and Cortana audio recordings, including sensitive conversations, without proper security measures or employee vetting. The lack of safeguards exposed users\u2019 private data to potential criminal or state access, constituting a significant privacy violation linked to AI system use.\n</incident>"}], "ground_truth": "data-exfiltration", "generated_content": "data-exfiltration"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n## Summary Component state manipulation is possible in `django-unicorn` due to missing access control checks during property updates and method calls. An attacker can bypass the intended `_is_public` protection to modify internal attributes such as `template_name` or trigger protected methods. ## Vulnerability Details: Component Access Control Bypass Security analysis identified that the framework fails to enforce visibility boundaries defined by `_is_public` within the action parsers. Specifically, the logic in `set_property_value()` and `_call_method_name()` utilizes `getattr` and `setattr` directly on component instances without verifying if the target attribute or method is explicitly marked as public. Vulnerability resides in: - `src/django_unicorn/views/action_parsers/call_method.py` - `src/django_unicorn/views/action_parsers/utils.py` While Django's template engine restricts rendering to registered directories, an unauthorized user can still force a component to render sensitive templates (e.g., admin layouts) from other installed applications or reset the component state by invoking the internal `reset()` method. ## Proof of Concept (PoC) Attacker can overwrite the `template_name` attribute by sending a crafted JSON payload to the message endpoint: 1. Construct a payload targeting a protected attribute: ```json { \"actionQueue\": [ { \"type\": \"syncInput\", \"payload\": { \"name\": \"template_name\", \"value\": \"admin/base.html\" } }\n</incident>"}], "ground_truth": "auth-bypass", "generated_content": "auth-bypass"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nFacebook&#x27;s and Twitter were not able to sufficiently moderate content of small language groups such as the Balkan languages using AI, allegedly due to the lack of investment in human moderation and difficulty in AI-solution design for the languages.\n</incident>"}], "ground_truth": "rce", "generated_content": "rce"}
-{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nA previously undisclosed facial recognition initiative operated by Israeli military intelligence units was reportedly deployed across Gaza after the October 7, 2023 attacks. According to multiple intelligence officers, the program uses Corsight technology alongside Google Photos to identify individuals from checkpoints, crowds, and drone footage. The system has allegedly produced misidentifications, including the widely reported detention of Palestinian poet Mosab Abu Toha on November 19, 2023.\n</incident>"}], "ground_truth": "rce", "generated_content": "rce"}
-{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nThe Australian government operates a voiceprint authentication system used by Centrelink and the Australian Taxation Office (ATO) that allows people to verify their identity over the phone using voice biometrics combined with customer reference numbers. As of February, 3.8 million Centrelink clients used the voiceprint service, and over 7.1 million people had verified their voice with the ATO. Guardian Australia conducted an investigation following reports of AI voice cloning attacks on overseas banking services. A Guardian journalist successfully created an AI clone of their own voice using just four minutes of audio and used this clone, combined with their customer reference number, to gain access to their own Centrelink self-service account. The voiceprint system is described by Services Australia as 'secure, accurate and reliable' and claims it can detect when someone is using a recording. However, the investigation demonstrated the vulnerability to AI-generated voice synthesis. The self-service phone system allows access to sensitive information including benefit payments and can authorize mailing of replacement cards. The technology is provided by Nuance, which claims its latest systems can detect synthetic voices in 86% to 99% of cases. When contacted about the vulnerability, Services Australia declined to say if the technology would be changed or removed.\n</incident>"}], "ground_truth": "data-exfiltration", "generated_content": "data-exfiltration"}
+{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nA previously undisclosed facial recognition initiative operated by Israeli military intelligence units was reportedly deployed across Gaza after the October 7, 2023 attacks. According to multiple intelligence officers, the program uses Corsight technology alongside Google Photos to identify individuals from checkpoints, crowds, and drone footage. The system has allegedly produced misidentifications, including the widely reported detention of Palestinian poet Mosab Abu Toha on November 19, 2023.\n</incident>"}], "ground_truth": "rce", "generated_content": "misinformation"}
+{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nThe Australian government operates a voiceprint authentication system used by Centrelink and the Australian Taxation Office (ATO) that allows people to verify their identity over the phone using voice biometrics combined with customer reference numbers. As of February, 3.8 million Centrelink clients used the voiceprint service, and over 7.1 million people had verified their voice with the ATO. Guardian Australia conducted an investigation following reports of AI voice cloning attacks on overseas banking services. A Guardian journalist successfully created an AI clone of their own voice using just four minutes of audio and used this clone, combined with their customer reference number, to gain access to their own Centrelink self-service account. The voiceprint system is described by Services Australia as 'secure, accurate and reliable' and claims it can detect when someone is using a recording. However, the investigation demonstrated the vulnerability to AI-generated voice synthesis. The self-service phone system allows access to sensitive information including benefit payments and can authorize mailing of replacement cards. The technology is provided by Nuance, which claims its latest systems can detect synthetic voices in 86% to 99% of cases. When contacted about the vulnerability, Services Australia declined to say if the technology would be changed or removed.\n</incident>"}], "ground_truth": "data-exfiltration", "generated_content": "deepfake"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n### Summary The `make:controller` CLI command calls `mkdir(..., recursive: true)` on a path built from the user-supplied controller name, **before** Nette's class-name validation runs. The class-file write is correctly rejected by Nette when the name contains `/`, but the recursive directory creation side effect is already committed \u2014 including directories located outside the project root through `../` traversal. ### Affected code `flight/commands/ControllerCommand.php` (\u2248 63-66): ```php if (is_dir(dirname($controllerPath)) === false) { $io->info('Creating directory ' . dirname($controllerPath), true); mkdir(dirname($controllerPath), 0755, true); // un-normalized, runs before validation } ``` ### Proof of concept ``` $ php vendor/flightphp/runway/runway make:controller '../../../../tmp/CONTROLLER_TRAVERSAL_TEST/pwn' Creating directory .../app/controllers/../../../../tmp/CONTROLLER_TRAVERSAL_TEST Nette\\InvalidArgumentException: Value '../../../../tmp/CONTROLLER_TRAVERSAL_TEST/pwnController' is not valid class name. $ ls /home/user/tmp/CONTROLLER_TRAVERSAL_TEST (directory exists \u2014 created before the exception was thrown) ``` ### Impact - **Arbitrary directory creation outside the project root**, executable by any local actor that can run the Flight CLI (developer machine, shared CI build agent, compromised dev container). - Primes log-file planting for chained LFI exploitation (e.g. creating a directory where an attacker can later drop a `.php` file to be include\n</incident>"}], "ground_truth": "path-traversal", "generated_content": "path-traversal"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nUS Congresswoman Alexandria Ocasio-Cortez became the target of AI-generated sexual deepfakes circulated on social media, drawing direct legislative action including her co-sponsored DEFIANCE Act creating civil remedies for non-consensual sexual deepfakes.\n</incident>"}], "ground_truth": "deepfake", "generated_content": "deepfake"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n## Summary Astro server actions have no default request body size limit, which can lead to memory exhaustion DoS. A single large POST to a valid action endpoint can crash the server process on memory-constrained deployments. ## Details On-demand rendered sites built with Astro can define server actions, which automatically parse incoming request bodies (JSON or FormData). The body is buffered entirely into memory with no size limit \u2014 a single oversized request is sufficient to exhaust the process heap and crash the server. Astro's Node adapter (`mode: 'standalone'`) creates an HTTP server with no body size protection. In containerized environments, the crashed process is automatically restarted, and repeated requests cause a persistent crash-restart loop. Action names are discoverable from HTML form attributes on any public page, so no authentication is required. ## PoC <details> ### Setup Create a new Astro project with the following files: `package.json`: ```json { \"name\": \"poc-dos\", \"private\": true, \"scripts\": { \"build\": \"astro build\", \"start:128mb\": \"node --max-old-space-size=128 dist/server/entry.mjs\" }, \"dependencies\": { \"astro\": \"5.17.2\", \"@astrojs/node\": \"9.5.3\" } } ``` `astro.config.mjs`: ```javascript import { defineConfig } from 'astro/config'; import node from '@astrojs/node'; export default defineConfig({ output: 'server', adapter: node({ mode: 'standalone' }), }); ``` `src/actions/index.ts`: ```typescript import { def\n</incident>"}], "ground_truth": "dos", "generated_content": "dos"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nAn AI-powered call center platform in the Middle East reportedly experienced a significant data breach, allegedly exposing over 10 million conversations between consumers, operators, and AI agents. Attackers allegedly accessed the platform\u2019s management dashboard, stealing sensitive data, including national ID documents. The breach poses reported risks such as phishing, identity theft, and social engineering attacks. The stolen data was reportedly listed for sale on the dark web.\n</incident>"}], "ground_truth": "data-exfiltration", "generated_content": "data-exfiltration"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n## Summary An attacker can cause a denial of service in webtransport-go by preventing or indefinitely delaying WebTransport session closure. A malicious peer can withhold QUIC flow control credit on the CONNECT stream, blocking transmission of the WT_CLOSE_SESSION capsule and causing the close operation to hang. ## Details WebTransport over HTTP/3 signals session termination by sending a WT_CLOSE_SESSION capsule on the CONNECT stream. The capsule is only needed to transmit a reason phrase and an error code to the peer. After the capsule is sent, the CONNECT stream is closed. In affected versions, the closure procedure blocked indefinitely while waiting for sufficient QUIC flow control credit from the peer. A malicious peer can withhold this credit, preventing the capsule from being sent. ## The Fix webtransport-go now attempts to send the WT_CLOSE_SESSION capsule with a short deadline. If the capsule cannot be sent within this deadline, the CONNECT stream is reset instead. This closes the WebTransport session promptly without transmitting the optional error details. This prevents indefinite blocking on session closure.\n</incident>"}], "ground_truth": "dos", "generated_content": "dos"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nThe fix for GHSA-p5g2-jm85-8g35 (ClickHouse SQL injection via aggregate query parameters) added column name validation to the `_aggregateBy` method but did not apply the same validation to three other query construction paths in `StatementGenerator`. The `toSortStatement`, `toSelectStatement`, and `toGroupByStatement` methods accept user-controlled object keys from API request bodies and interpolate them as ClickHouse `Identifier` parameters without verifying they correspond to actual model columns. ClickHouse Identifier parameters are substituted directly into queries without escaping, so an attacker who can reach any analytics list or aggregate endpoint can inject arbitrary SQL through crafted `sort`, `select`, or `groupBy` keys. ## Details ### Root cause `StatementGenerator.ts` has four methods that iterate over user-provided object keys to build SQL: | Method | Validates keys? | |--------|----------------| | `toWhereStatement` (line 292) | Yes - calls `this.model.getTableColumn(key)` | | `toSortStatement` (line 467) | **No** | | `toSelectStatement` (line 483) | **No** | | `toGroupByStatement` (line 451) | **No** | In `Statement.ts`, when a value passed to the `SQL` tagged template is a string, it receives the `Identifier` data type (line 40). Per [ClickHouse documentation](https://clickhouse.com/docs/en/sql-reference/syntax#defining-and-using-query-parameters), Identifier parameters are substituted directly into the query without quoting or escaping. This is correct\n</incident>"}], "ground_truth": "deserialization", "generated_content": "data-exfiltration"}
-{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n### Impact picomatch is vulnerable to a **method injection vulnerability (CWE-1321)** affecting the `POSIX_REGEX_SOURCE` object. Because the object inherits from `Object.prototype`, specially crafted POSIX bracket expressions (e.g., `[[:constructor:]]`) can reference inherited method names. These methods are implicitly converted to strings and injected into the generated regular expression. This leads to **incorrect glob matching behavior (integrity impact)**, where patterns may match unintended filenames. The issue does **not enable remote code execution**, but it can cause security-relevant logic errors in applications that rely on glob matching for filtering, validation, or access control. All users of affected `picomatch` versions that process untrusted or user-controlled glob patterns are potentially impacted. ### Patches This issue is fixed in picomatch 4.0.4, 3.0.2 and 2.3.2. Users should upgrade to one of these versions or later, depending on their supported release line. ### Workarounds If upgrading is not immediately possible, avoid passing untrusted glob patterns to picomatch. Possible mitigations include: - Sanitizing or rejecting untrusted glob patterns, especially those containing POSIX character classes like `[[:...:]]`. - Avoiding the use of POSIX bracket expressions if user input is involved. - Manually patching the library by modifying `POSIX_REGEX_SOURCE` to use a null prototype: ```js const POSIX_REGEX_SOURCE = { __proto__: null, alnum\n</incident>"}], "ground_truth": "auth-bypass", "generated_content": "path-traversal"}
+{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n### Impact picomatch is vulnerable to a **method injection vulnerability (CWE-1321)** affecting the `POSIX_REGEX_SOURCE` object. Because the object inherits from `Object.prototype`, specially crafted POSIX bracket expressions (e.g., `[[:constructor:]]`) can reference inherited method names. These methods are implicitly converted to strings and injected into the generated regular expression. This leads to **incorrect glob matching behavior (integrity impact)**, where patterns may match unintended filenames. The issue does **not enable remote code execution**, but it can cause security-relevant logic errors in applications that rely on glob matching for filtering, validation, or access control. All users of affected `picomatch` versions that process untrusted or user-controlled glob patterns are potentially impacted. ### Patches This issue is fixed in picomatch 4.0.4, 3.0.2 and 2.3.2. Users should upgrade to one of these versions or later, depending on their supported release line. ### Workarounds If upgrading is not immediately possible, avoid passing untrusted glob patterns to picomatch. Possible mitigations include: - Sanitizing or rejecting untrusted glob patterns, especially those containing POSIX character classes like `[[:...:]]`. - Avoiding the use of POSIX bracket expressions if user input is involved. - Manually patching the library by modifying `POSIX_REGEX_SOURCE` to use a null prototype: ```js const POSIX_REGEX_SOURCE = { __proto__: null, alnum\n</incident>"}], "ground_truth": "auth-bypass", "generated_content": "command-injection"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n### Summary A denial-of-service vulnerability exists in the SM2 public-key encryption (PKE) implementation: the `decrypt()` path performs unchecked `slice::split_at` operations on input buffers derived from untrusted ciphertext. An attacker can submit short/undersized ciphertext or carefully-crafted DER-encoded structures to trigger bounds-check panics (Rust unwinding) which crash the calling thread or process. ### Affected Component / Versions - File: `src/pke/decrypting.rs` - Functions: `DecryptingKey::decrypt_digest/decrypt/decrypt_der`, internal `decrypt()` implementation - Affected releases: - sm2 0.14.0-rc.0 (https://crates.io/crates/sm2/0.14.0-rc.0) - sm2 0.14.0-pre.0 (https://crates.io/crates/sm2/0.14.0-pre.0) ### Details The vulnerability is located in the file `sm2/src/pke/decrypting.rs`. The **fundamental cause** of the vulnerability is that the decryption function **does not strictly check** the ciphertext's format and length information. Consequently, a maliciously crafted ciphertext can trigger Rust's **panic mechanism** instead of the expected error handling (`Error`) mechanism. The Rust function `C.split_at(L)` will trigger a Panic if the length is less than `L`, as shown in the code comment below: the `decrypting` function has **at least three locations** where a slice operation might trigger a Panic. ```rust fn decrypt( secret_scalar: &Scalar, mode: Mode, hasher: &mut dyn DynDigest, cipher: &[u8], ) -> Result<Vec<u8>>\n</incident>"}], "ground_truth": "dos", "generated_content": "dos"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n### Summary A path traversal vulnerability in `IArchive.WriteToDirectory()` allows a malicious archive to create directories outside the intended extraction root. For TAR archives, this can be escalated to arbitrary file writes by chaining with a symlink entry, giving a full write primitive on the target filesystem subject to the permissions of the running process. ### Details The vulnerable code is in the directory-entry branch of `WriteToDirectoryInternal` (sync, `IArchiveExtensions.cs:48\u201361`) and `WriteToDirectoryAsyncInternal` (async, `IAsyncArchiveExtensions.cs:70\u201384`): ```csharp var dirPath = Path.Combine(destinationDirectory, entry.Key); Directory.CreateDirectory(Path.GetDirectoryName(dirPath + \"/\")); ``` No `Path.GetFullPath()` normalisation and no bounds check are applied before the `Directory.CreateDirectory` call. Two .NET `Path.Combine` behaviours make this exploitable: - **Relative traversal**: `Path.Combine(\"/safe/extract\", \"../../evil\")` \u2192 the OS resolves `..` segments on the raw path, placing the directory outside the extraction root. - **Absolute path override**: `Path.Combine(\"/safe/extract\", \"/tmp/evil\")` \u2192 returns `\"/tmp/evil\"` \u2014 the base is discarded entirely for rooted paths. File entries are **not** directly affected \u2014 they route through `ExtractionMethods.WriteEntryToDirectory` which applies the correct guard (`GetFullPath` + `StartsWith`, see `ExtractionMethods.cs:54\u201365`). The directory-entry branch is a separate fast-path that was added withou\n</incident>"}], "ground_truth": "path-traversal", "generated_content": "path-traversal"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\n## Summary The SignalK Server exposes an unauthenticated HTTP endpoint that allows remote attackers to modify navigation data source priorities. This endpoint, accessible via `PUT /signalk/v1/api/sourcePriorities`, does not enforce authentication or authorization checks and directly assigns user-controlled input to the server configuration. As a result, attackers can influence which GPS, AIS, or other sensor data sources are trusted by the system. The changes are immediately applied and persisted to disk, allowing the manipulation to survive server restarts. ### Affected Component - **File**: `src/serverroutes.ts` - **Endpoint**: `PUT /signalk/v1/api/sourcePriorities` (also accessible at `/skServer/sourcePriorities`) - **Lines**: 1064-1076 - **Function**: Source priorities configuration handler ### Vulnerable Code ```typescript // src/serverroutes.ts - Lines 1064-1076 app.put( `${SERVERROUTESPREFIX}/sourcePriorities`, (req: Request, res: Response) => { app.config.settings.sourcePriorities = req.body app.activateSourcePriorities() writeSettingsFile(app, app.config.settings, (err: any) => { if (err) { res .status(500) .send('Unable to save to sourcePrefences in settings file') } else { res.json({ result: 'ok' }) } }) } ) ``` ## Vulnerability Characteristics **Missing Authentication**: The endpoint has zero authentication middleware, allowing unauthenticated access from any network-adjacent attacke\n</incident>"}], "ground_truth": "auth-bypass", "generated_content": "auth-bypass"}
@@ -216,7 +216,7 @@
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nMagento Open Source versions 2.4.7-p2, 2.4.6-p7, 2.4.5-p9, 2.4.4-p10 and earlier are affected by an Information Exposure vulnerability that could result in a security feature bypass. An admin attacker could leverage this vulnerability to have a low impact on confidentiality which may aid in further attacks. Exploitation of this issue does not require user interaction.\n</incident>"}], "ground_truth": "info-disclosure", "generated_content": "info-disclosure"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nAIAAIC report: Paran\u00e1 school attendance facial recognition system criticised as \"invasive\", \"noxious\". System: LRCO Paran\u00e1. Technology: Facial recognition. Purpose: Register student attendance. Ethical issues: Accountability; Accuracy/reliability; Consent; Fairness; Normalisation; Power inbalance; Privacy/surveillance. Reported consequences: Litigation; Legislative proposal.\n</incident>"}], "ground_truth": "privacy-violation", "generated_content": "privacy-violation"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nByteDance allegedly scraped short-form videos, usernames, profile pictures, and descriptions of accounts on Instagram, Snapchat, and other sources, and uploaded them without consent on Flipagram, TikTok\u2019s predecessor, in order to improve its \u201cFor You\u201d algorithm&#x27;s performance on American users.\n</incident>"}], "ground_truth": "rce", "generated_content": "data-exfiltration"}
-{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nMeta AI launched Galactica \u2014 a large language model trained on scientific literature and designed to assist with scientific writing, summarisation, and knowledge retrieval \u2014 publicly via a demo on November 15, 2022. Within 72 hours, Meta withdrew the public demo after widespread criticism from the scientific community. Researchers found that Galactica confidently generated plausible-sounding but factually incorrect scientific text, including fabricated citations, incorrect chemical formulae, and authoritative-sounding passages on controversial topics (e.g., the history of bears in space). The core failure was that the model's confident, authoritative tone conveyed false certainty \u2014 users with limited domain expertise could not distinguish accurate from fabricated content. This remains the canonical real-world example of LLM misinformation at launch in a high-stakes domain.\n</incident>"}], "ground_truth": "misinformation", "generated_content": "misinformation"}
+{"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nMeta AI launched Galactica \u2014 a large language model trained on scientific literature and designed to assist with scientific writing, summarisation, and knowledge retrieval \u2014 publicly via a demo on November 15, 2022. Within 72 hours, Meta withdrew the public demo after widespread criticism from the scientific community. Researchers found that Galactica confidently generated plausible-sounding but factually incorrect scientific text, including fabricated citations, incorrect chemical formulae, and authoritative-sounding passages on controversial topics (e.g., the history of bears in space). The core failure was that the model's confident, authoritative tone conveyed false certainty \u2014 users with limited domain expertise could not distinguish accurate from fabricated content. This remains the canonical real-world example of LLM misinformation at launch in a high-stakes domain.\n</incident>"}], "ground_truth": "misinformation", "generated_content": "rce"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nA flaw has been found in Ollama up to 18.1. This issue affects some unknown processing of the file server/download.go of the component Model Pull API. Executing a manipulation can lead to server-side request forgery. The attack can be launched remotely. The vendor was contacted early about this disclosure but did not respond in any way.\n</incident>"}], "ground_truth": "ssrf", "generated_content": "ssrf"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nAgentAPI is an HTTP API for Claude Code, Goose, Aider, Gemini, Amp, and Codex. Versions 0.3.3 and below are susceptible to a client-side DNS rebinding attack when hosted over plain HTTP on localhost. An attacker can gain access to the /messages endpoint served by the Agent API. This allows for the unauthorized exfiltration of sensitive user data, specifically local message history, which can include secret keys, file system contents, and intellectual property the user was working on locally. This issue is fixed in version 0.4.0.\n</incident>"}], "ground_truth": "data-exfiltration", "generated_content": "data-exfiltration"}
 {"messages": [{"role": "user", "content": "What attack vector does the following GenAI security incident describe?\n\n<incident>\nClearview AI, known for its facial recognition system with a database of over 3 billion photos, suffered a data breach exposing its entire client list, including law enforcement agencies, and usage details. The incident raised significant privacy concerns and prompted criticism from lawmakers and investigations by privacy authorities.\n</incident>"}], "ground_truth": "data-exfiltration", "generated_content": "data-exfiltration"}

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:da3e7916a9c3a9bdc1d65a3a86e68c6c06ab04fd280e02675e7357542392c8fa
+oid sha256:b5aae0a2160af9bfafbf734dc2e956a27d2ec4ded9101f3494b0466f6c3c55ff
 size 5496