Initial release: DATABASE_v9_1 mirror from GitHub source-of-truth

DATABASE_v9_1.csv

@@ -0,0 +1,13 @@
+entry_id,family,size,arch,n_q,n_kv,qk_norm,training_tokens,eta_peak,lambda_wd,T_steps,tau_iter,T_over_tau,Physical_State,hp_confidence,n_records,k_median_qkv,lambda_median_qkv,R2_median_qkv,R2_below_95_qkv,k_median_o,lambda_median_o,R2_median_o,R2_below_95_o,k_median_ffn_in,lambda_median_ffn_in,R2_median_ffn_in,R2_below_95_ffn_in,k_median_ffn_out,lambda_median_ffn_out,R2_median_ffn_out,R2_below_95_ffn_out,k_median_q,lambda_median_q,R2_median_q,R2_below_95_q,k_median_k,lambda_median_k,R2_median_k,R2_below_95_k,k_median_v,lambda_median_v,R2_median_v,R2_below_95_v,k_median_gate,lambda_median_gate,R2_median_gate,R2_below_95_gate,k_median_up,lambda_median_up,R2_median_up,R2_below_95_up,k_median_down,lambda_median_down,R2_median_down,R2_below_95_down
+pythia-70m,Pythia,70m,MHA-merged,,,False,300B,0.001,0.01,143000,100000,1.43,Saturated,explicit,24,1.0488499908322977,0.027829200076895988,0.9980661795697541,1/6,1.1837935286072323,0.018602428920403802,0.9985356796636042,0/6,1.1902710504479086,0.026003913070232375,0.9983539238333419,0/6,1.1898015122737948,0.021012236928099785,0.9982397335063831,0/6,,,,,,,,,,,,,,,,,,,,,,,,
+pythia-160m,Pythia,160m,MHA-merged,,,False,300B,0.0006,0.01,143000,166667,0.858,Near-saturated,explicit,48,1.0982363830376656,0.02515902104056176,0.9994532449901053,0/12,1.191740628409264,0.015679046523988435,0.9982396364419186,0/12,1.1926538220592602,0.024112364246690256,0.9982122944417534,0/12,1.1953319289698738,0.018883547673911608,0.9981279582557228,0/12,,,,,,,,,,,,,,,,,,,,,,,,
+pythia-410m,Pythia,410m,MHA-merged,,,False,300B,0.0003,0.01,143000,333333,0.429,Approaching,explicit,96,1.146561528599217,0.02122171415636209,0.998936826903025,0/24,1.1988129781844883,0.017971449756907294,0.9980865208585606,0/24,1.1990799021662437,0.020800650303664674,0.9979860387362547,0/24,1.201096993470142,0.01715614873199387,0.9979236660582246,0/24,,,,,,,,,,,,,,,,,,,,,,,,
+pythia-1b,Pythia,1B,MHA-merged,,,False,300B,0.0003,0.01,143000,333333,0.429,Approaching,explicit,64,1.1758165081776015,0.020136835465006095,0.9986403658408189,0/16,1.199000460660631,0.018768755040468516,0.9980321219869854,0/16,1.199710534194005,0.020151192747180725,0.9980168719428919,0/16,1.1992630782799794,0.01733192000700065,0.9980239716959693,0/16,,,,,,,,,,,,,,,,,,,,,,,,
+pythia-6.9b,Pythia,6.9B,MHA-merged,,,False,300B,0.00012,0.01,143000,833333,0.172,Transition,explicit,128,1.1725961308622437,0.016787633359452464,0.9989545055530709,0/32,1.1986053979939104,0.013017965092167658,0.998024607412637,0/32,1.2026301500842531,0.016750572288110506,0.9978736032238523,0/32,1.200956915459356,0.014671536899830185,0.9979452059317363,0/32,,,,,,,,,,,,,,,,,,,,,,,,
+olmo-1-7b,OLMo-1,7B,MHA-separate,32,32,False,2.5T,0.0003,0.1,477000,33333,14.31,Saturated,explicit,224,,,,,1.0408657276990618,0.002401554550633505,0.9971792835186756,2/32,,,,,,,,,0.8122912154036579,0.001728964705557532,0.9987052449748253,12/32,0.760128189603047,0.0017030071063091484,0.9985575728031171,13/32,1.0601198634820659,0.0025129714668301892,0.9971448334283877,0/32,1.2009840121871695,0.003363367499645852,0.997926463922367,0/32,1.203942863882499,0.0034419612928361385,0.9977918785255582,0/32,1.204103938938425,0.0034857079668821657,0.9977980830738356,0/32
+olmo-2-7b,OLMo-2,7B,MHA-separate,32,32,True,5T,0.0003,0.1,600000,33333,18.0,Saturated,inferred,224,,,,,1.1957824956841125,0.01633689288553169,0.9980401478865495,0/32,,,,,,,,,0.989513585691632,0.014060155205477352,0.9908743229136651,2/32,0.9715995155742865,0.013240338570321912,0.9935710555725896,7/32,1.1930013229215812,0.016630755274009916,0.9980887579980875,0/32,1.1976133728932,0.017549626600536647,0.9980961344940342,0/32,1.2031829219410521,0.01675671031075271,0.9978851898208778,0/32,1.2030757372152108,0.01654031946425652,0.997865366465111,0/32
+llama-3-8b,Llama-3,8B,GQA-4:1,32,8,False,15T,0.0003,0.1,1000000,33333,30.0,Saturated,inferred,224,,,,,1.184071385746421,0.008293435300471455,0.998465712933232,0/32,,,,,,,,,1.1352115988550961,0.013524263847823039,0.9995330340960915,0/32,1.1461534355136989,0.0210008606786853,0.9994212651102061,0/32,1.1710054472835778,0.0074532770403264465,0.9984796629483923,0/32,1.1890150725270956,0.011714635540860925,0.9984570553256354,0/32,1.1970808681863536,0.009321699181305845,0.9979506822747821,0/32,1.1931330108175051,0.009227488024057286,0.9980590165512454,0/32
+mistral-7b,Mistral,7B,GQA-4:1,32,8,False,8T,0.0003,0.1,500000,33333,15.0,Saturated,estimated,224,,,,,1.190243912925892,0.002402403275884374,0.9983281530362517,0/32,,,,,,,,,1.1485392032119919,0.0029794427980022876,0.999306901399193,1/32,1.1290637579544684,0.0035066491246891554,0.9996222872848544,0/32,1.1701993384534188,0.0023825545163240107,0.9983546533501352,0/32,1.194720188713138,0.0028355399128622725,0.9981770951518927,0/32,1.1964326908213916,0.002581969834286769,0.9980098695080295,0/32,1.1925738330014424,0.002545509262438919,0.9981714252724023,0/32
+qwen2.5-7b,Qwen2.5,7B,GQA-7:1,28,4,False,18T,0.0003,0.1,1100000,33333,33.0,Saturated,inferred,196,,,,,1.1664607878011877,0.013070396382489802,0.9989430337848695,0/28,,,,,,,,,1.1328116652061784,0.013520394365978409,0.9993364777176525,0/28,1.1033345281197386,0.014088124992005789,0.9998005555777307,0/28,1.1430042518328158,0.01389958738996697,0.9986895435422882,0/28,1.190394620384456,0.01430298685422907,0.9982928398948381,0/28,1.1888191751907358,0.014378046711090012,0.9982963964287199,0/28,1.1829726136894823,0.014066967405546603,0.9985022496848237,0/28
+qwen2.5-14b,Qwen2.5,14B,GQA-5:1,40,8,False,18T,0.0003,0.1,1100000,33333,33.0,Saturated,estimated,189,,,,,1.1841492889945942,0.01635424142703244,0.9985449642031292,0/27,,,,,,,,,1.1598333532217517,0.017467613550818994,0.9988577504081926,0/27,1.134601112331773,0.019688051072659123,0.9993206828179739,0/27,1.1636461389324964,0.016113911185781107,0.9985097429190216,0/27,1.1908688512686276,0.01786358668592591,0.9982032807344513,3/27,1.1914319791119936,0.018228218923577316,0.9980806394042295,1/27,1.1884973746873577,0.018154542924089155,0.9982328415038318,1/27
+qwen3-8b,Qwen3,8B,GQA-4:1,32,8,True,36T,0.0003,0.1,2200000,33333,66.0,Saturated,inferred,252,,,,,1.1801768316123442,0.022653566194551916,0.9985957983960119,0/36,,,,,,,,,1.162275174612215,0.02149767664699718,0.9991031504379486,0/36,1.1539186367719623,0.02110274935915082,0.9992249255351853,0/36,1.1580956333123513,0.024755221044291426,0.9986119730450285,0/36,1.1872130646195065,0.02374112314986921,0.998329750229662,0/36,1.1885651746494985,0.023673103492721705,0.9981567875217829,1/36,1.1846459565046368,0.023230312440040608,0.9984190108999799,0/36

DATABASE_v9_1.md

@@ -0,0 +1,70 @@
+# DATABASE_v9_1 — Reference Table (12 family entries)
+
+**Source**: cascade v3 (per-component Weibull fit, mid-80% trim)
+**Generated**: by populate_database_v9_1.py
+
+## 12 entries (5 Pythia size + 7 cross-family) — k median per component
+
+| Entry | Family | Size | Architecture | n_q/n_kv | QK-Norm | tokens | k_q | k_k | k_v | k_o | k_gate | k_up | k_down | low-R²(q) |
+|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+| pythia-70m | Pythia | 70m | MHA-merged | merged | no | 300B | 1.0488 | — | — | 1.1838 | 0.0000 | 0.0000 | 0.0000 | 1/6 |
+| pythia-160m | Pythia | 160m | MHA-merged | merged | no | 300B | 1.0982 | — | — | 1.1917 | 0.0000 | 0.0000 | 0.0000 | 0/12 |
+| pythia-410m | Pythia | 410m | MHA-merged | merged | no | 300B | 1.1466 | — | — | 1.1988 | 0.0000 | 0.0000 | 0.0000 | 0/24 |
+| pythia-1b | Pythia | 1B | MHA-merged | merged | no | 300B | 1.1758 | — | — | 1.1990 | 0.0000 | 0.0000 | 0.0000 | 0/16 |
+| pythia-6.9b | Pythia | 6.9B | MHA-merged | merged | no | 300B | 1.1726 | — | — | 1.1986 | 0.0000 | 0.0000 | 0.0000 | 0/32 |
+| olmo-1-7b | OLMo-1 | 7B | MHA-separate | 32/32 | no | 2.5T | 0.8123 | 0.7601 | 1.0601 | 1.0409 | 1.2010 | 1.2039 | 1.2041 | 12/32 |
+| olmo-2-7b | OLMo-2 | 7B | MHA-separate | 32/32 | yes | 5T | 0.9895 | 0.9716 | 1.1930 | 1.1958 | 1.1976 | 1.2032 | 1.2031 | 2/32 |
+| llama-3-8b | Llama-3 | 8B | GQA-4:1 | 32/8 | no | 15T | 1.1352 | 1.1462 | 1.1710 | 1.1841 | 1.1890 | 1.1971 | 1.1931 | 0/32 |
+| mistral-7b | Mistral | 7B | GQA-4:1 | 32/8 | no | 8T | 1.1485 | 1.1291 | 1.1702 | 1.1902 | 1.1947 | 1.1964 | 1.1926 | 1/32 |
+| qwen2.5-7b | Qwen2.5 | 7B | GQA-7:1 | 28/4 | no | 18T | 1.1328 | 1.1033 | 1.1430 | 1.1665 | 1.1904 | 1.1888 | 1.1830 | 0/28 |
+| qwen2.5-14b | Qwen2.5 | 14B | GQA-5:1 | 40/8 | no | 18T | 1.1598 | 1.1346 | 1.1636 | 1.1841 | 1.1909 | 1.1914 | 1.1885 | 0/27 |
+| qwen3-8b | Qwen3 | 8B | GQA-4:1 | 32/8 | yes | 36T | 1.1623 | 1.1539 | 1.1581 | 1.1802 | 1.1872 | 1.1886 | 1.1846 | 0/36 |
+
+## Training hyperparameters + T/τ + Physical State (Wang-Aitchison 2024 cycle ratio)
+
+τ_iter = 1/(η · λ_wd) — EMA iteration time-constant.  T/τ_iter = T_steps / τ_iter — completed EMA cycles.
+
+| Entry | η_peak | λ_wd | T_steps | τ_iter | **T/τ** | **Physical State** | hp source |
+|---|---|---|---|---|---|---|---|
+| pythia-70m | 1.0e-03 | 0.01 | 143000 | 100000 | **1.43** | **Saturated** | explicit |
+| pythia-160m | 6.0e-04 | 0.01 | 143000 | 166667 | **0.86** | **Near-saturated** | explicit |
+| pythia-410m | 3.0e-04 | 0.01 | 143000 | 333333 | **0.43** | **Approaching** | explicit |
+| pythia-1b | 3.0e-04 | 0.01 | 143000 | 333333 | **0.43** | **Approaching** | explicit |
+| pythia-6.9b | 1.2e-04 | 0.01 | 143000 | 833333 | **0.17** | **Transition** | explicit |
+| olmo-1-7b | 3.0e-04 | 0.1 | 477000 | 33333 | **14.31** | **Saturated** | explicit |
+| olmo-2-7b | 3.0e-04 | 0.1 | 600000 | 33333 | **18.00** | **Saturated** | inferred |
+| llama-3-8b | 3.0e-04 | 0.1 | 1000000 | 33333 | **30.00** | **Saturated** | inferred |
+| mistral-7b | 3.0e-04 | 0.1 | 500000 | 33333 | **15.00** | **Saturated** | estimated |
+| qwen2.5-7b | 3.0e-04 | 0.1 | 1100000 | 33333 | **33.00** | **Saturated** | inferred |
+| qwen2.5-14b | 3.0e-04 | 0.1 | 1100000 | 33333 | **33.00** | **Saturated** | estimated |
+| qwen3-8b | 3.0e-04 | 0.1 | 2200000 | 33333 | **66.00** | **Saturated** | inferred |
+
+**Physical State thresholds** (Wang-Aitchison 2024 cycle ratio):
+
+- Saturated: T/τ ≥ 1.20
+- Near-saturated: 0.80 ≤ T/τ < 1.20
+- Approaching: 0.40 ≤ T/τ < 0.80
+- Partial: 0.25 ≤ T/τ < 0.40
+- Transition: T/τ < 0.25
+
+**hp source confidence**:
+
+- *explicit*: paper Table / official tech report directly states the value
+- *inferred*: paper §3 states a quantity from which we derive it (e.g. tokens × batch / seq → steps)
+- *estimated*: paper does not publish; same-family typical recipe used as fallback
+
+---
+
+## Verification
+
+Per-entry per-component sanity check is recorded in
+[`DATABASE_v9_1_report.md`](DATABASE_v9_1_report.md). All entries pass
+R² ≥ 0.99 on the Transmission Class components.
+
+**Transmission Class aggregated band** (median across components per
+entry, then aggregated across the 12 entries): k ∈ [1.186, 1.204],
+cross-family CV = 0.51%. See paper §3 for the strict-band definition
+and protocol.
+
+For per-block raw fits and the cascade pipeline that produces this
+table, see the `npm-weibull-py` repository on GitHub.

DATABASE_v9_1_report.md

@@ -0,0 +1,82 @@
+# DATABASE_v9_1 sanity verify report
+
+### pythia-70m (Pythia 70m, MHA-merged)
+  records: 24; tokens: 300B; QK-Norm: False
+  k_median(qkv) = 1.0488, R2(qkv) = 0.9981, low-R2: 1/6
+  k_median(gate)=n/a, k_median(up)=n/a, k_median(down)=n/a
+
+### pythia-160m (Pythia 160m, MHA-merged)
+  records: 48; tokens: 300B; QK-Norm: False
+  k_median(qkv) = 1.0982, R2(qkv) = 0.9995, low-R2: 0/12
+  k_median(gate)=n/a, k_median(up)=n/a, k_median(down)=n/a
+
+### pythia-410m (Pythia 410m, MHA-merged)
+  records: 96; tokens: 300B; QK-Norm: False
+  k_median(qkv) = 1.1466, R2(qkv) = 0.9989, low-R2: 0/24
+  k_median(gate)=n/a, k_median(up)=n/a, k_median(down)=n/a
+
+### pythia-1b (Pythia 1B, MHA-merged)
+  records: 64; tokens: 300B; QK-Norm: False
+  k_median(qkv) = 1.1758, R2(qkv) = 0.9986, low-R2: 0/16
+  k_median(gate)=n/a, k_median(up)=n/a, k_median(down)=n/a
+
+### pythia-6.9b (Pythia 6.9B, MHA-merged)
+  records: 128; tokens: 300B; QK-Norm: False
+  k_median(qkv) = 1.1726, R2(qkv) = 0.9990, low-R2: 0/32
+  k_median(gate)=n/a, k_median(up)=n/a, k_median(down)=n/a
+
+### olmo-1-7b (OLMo-1 7B, MHA-separate)
+  records: 224; tokens: 2.5T; QK-Norm: False
+  k_median(q)  = 0.8123, R2(q)  = 0.9987, low-R2: 12/32
+  k_median(k)  = 0.7601, R2(k)  = 0.9986, low-R2: 13/32
+  k_median(v)  = 1.0601, R2(v)  = 0.9971
+  k_median(o)  = 1.0409, R2(o)  = 0.9972
+  k_median(gate)=1.2010, k_median(up)=1.2039, k_median(down)=1.2041
+
+### olmo-2-7b (OLMo-2 7B, MHA-separate)
+  records: 224; tokens: 5T; QK-Norm: True
+  k_median(q)  = 0.9895, R2(q)  = 0.9909, low-R2: 2/32
+  k_median(k)  = 0.9716, R2(k)  = 0.9936, low-R2: 7/32
+  k_median(v)  = 1.1930, R2(v)  = 0.9981
+  k_median(o)  = 1.1958, R2(o)  = 0.9980
+  k_median(gate)=1.1976, k_median(up)=1.2032, k_median(down)=1.2031
+
+### llama-3-8b (Llama-3 8B, GQA-4:1)
+  records: 224; tokens: 15T; QK-Norm: False
+  k_median(q)  = 1.1352, R2(q)  = 0.9995, low-R2: 0/32
+  k_median(k)  = 1.1462, R2(k)  = 0.9994, low-R2: 0/32
+  k_median(v)  = 1.1710, R2(v)  = 0.9985
+  k_median(o)  = 1.1841, R2(o)  = 0.9985
+  k_median(gate)=1.1890, k_median(up)=1.1971, k_median(down)=1.1931
+
+### mistral-7b (Mistral 7B, GQA-4:1)
+  records: 224; tokens: 8T; QK-Norm: False
+  k_median(q)  = 1.1485, R2(q)  = 0.9993, low-R2: 1/32
+  k_median(k)  = 1.1291, R2(k)  = 0.9996, low-R2: 0/32
+  k_median(v)  = 1.1702, R2(v)  = 0.9984
+  k_median(o)  = 1.1902, R2(o)  = 0.9983
+  k_median(gate)=1.1947, k_median(up)=1.1964, k_median(down)=1.1926
+
+### qwen2.5-7b (Qwen2.5 7B, GQA-7:1)
+  records: 196; tokens: 18T; QK-Norm: False
+  k_median(q)  = 1.1328, R2(q)  = 0.9993, low-R2: 0/28
+  k_median(k)  = 1.1033, R2(k)  = 0.9998, low-R2: 0/28
+  k_median(v)  = 1.1430, R2(v)  = 0.9987
+  k_median(o)  = 1.1665, R2(o)  = 0.9989
+  k_median(gate)=1.1904, k_median(up)=1.1888, k_median(down)=1.1830
+
+### qwen2.5-14b (Qwen2.5 14B, GQA-5:1)
+  records: 189; tokens: 18T; QK-Norm: False
+  k_median(q)  = 1.1598, R2(q)  = 0.9989, low-R2: 0/27
+  k_median(k)  = 1.1346, R2(k)  = 0.9993, low-R2: 0/27
+  k_median(v)  = 1.1636, R2(v)  = 0.9985
+  k_median(o)  = 1.1841, R2(o)  = 0.9985
+  k_median(gate)=1.1909, k_median(up)=1.1914, k_median(down)=1.1885
+
+### qwen3-8b (Qwen3 8B, GQA-4:1)
+  records: 252; tokens: 36T; QK-Norm: True
+  k_median(q)  = 1.1623, R2(q)  = 0.9991, low-R2: 0/36
+  k_median(k)  = 1.1539, R2(k)  = 0.9992, low-R2: 0/36
+  k_median(v)  = 1.1581, R2(v)  = 0.9986
+  k_median(o)  = 1.1802, R2(o)  = 0.9986
+  k_median(gate)=1.1872, k_median(up)=1.1886, k_median(down)=1.1846

LICENSE

@@ -0,0 +1,32 @@
+Creative Commons Attribution 4.0 International (CC BY 4.0)
+
+Copyright (c) 2026 Tiexin Ding
+
+Full license text: https://creativecommons.org/licenses/by/4.0/legalcode
+
+You are free to:
+
+  Share — copy and redistribute the material in any medium or format
+  Adapt — remix, transform, and build upon the material for any
+          purpose, even commercially.
+
+Under the following terms:
+
+  Attribution — You must give appropriate credit, provide a link to
+                the license, and indicate if changes were made. You
+                may do so in any reasonable manner, but not in any
+                way that suggests the licensor endorses you or your
+                use.
+
+  No additional restrictions — You may not apply legal terms or
+                technological measures that legally restrict others
+                from doing anything the license permits.
+
+The licensor cannot revoke these freedoms as long as you follow the
+license terms.
+
+When citing this dataset, please cite the associated paper:
+
+  Ding, Tiexin (2026). A Two-Parameter Weibull Framework for
+  Diagnosing Transformer Weight Distributions.
+  arXiv:2605.18898. https://doi.org/10.48550/arXiv.2605.18898

README.md

@@ -0,0 +1,160 @@
+---
+license: cc-by-4.0
+task_categories:
+  - other
+language:
+  - en
+size_categories:
+  - n<1K
+tags:
+  - weibull
+  - transformer
+  - weight-statistics
+  - model-diagnostics
+  - npm-weibull
+pretty_name: NPM-Weibull DATABASE v9_1
+configs:
+  - config_name: default
+    data_files: DATABASE_v9_1.csv
+---
+
+# NPM-Weibull DATABASE v9_1
+
+Benchmark database of Weibull (k, λ) fits for **12 transformer model entries** spanning **7 architectural families** (Pythia, OLMo-1/2, LLaMA-3, Mistral, Qwen2.5, Qwen3) — 70M–14B parameters, GeLU and SwiGLU activations, Pre-LN and QK-Norm placements.
+
+- 📄 **Paper**: [A Two-Parameter Weibull Framework for Diagnosing Transformer Weight Distributions](https://arxiv.org/abs/2605.18898) (arXiv:2605.18898)
+- 📦 **Source repo (primary)**: [github.com/tiexinding/NPM-Weibull-public](https://github.com/tiexinding/NPM-Weibull-public) — directory `database_v9_1/`
+- 🔧 **Python library**: `pip install npm-weibull-py` — [PyPI](https://pypi.org/project/npm-weibull-py/)
+
+---
+
+## 📌 Source of truth
+
+The **primary, authoritative location** of this dataset is the GitHub repository above (directory `database_v9_1/`). This Hugging Face Dataset is a **mirror** synced from GitHub on each tagged release, provided for `load_dataset(...)` convenience and Dataset Viewer discoverability.
+
+- Updates are **batched per release**, not per commit — for the latest state, watch the GitHub repo.
+- Issues and pull requests: please file on [GitHub Issues](https://github.com/tiexinding/NPM-Weibull-public/issues) rather than HF discussions (faster response).
+- License, citation, and documentation are kept consistent across both surfaces.
+
+---
+
+## Quick start
+
+### Option 1 — `load_dataset` (Hugging Face)
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("TiexinDing/NPM-Weibull-DATABASE-v9_1")
+print(ds["train"][0])  # First entry: pythia-70m
+```
+
+### Option 2 — pandas
+
+```python
+import pandas as pd
+
+df = pd.read_csv("DATABASE_v9_1.csv")
+print(df[["entry_id", "k_median_o", "lambda_median_o"]].head())
+```
+
+### Option 3 — `npm-weibull-py` library (recommended for analysis)
+
+```python
+from npm_weibull import DATABASE_v9_1, compare_to_benchmark
+
+print(DATABASE_v9_1)  # 12 entries with per-component fits
+
+cmp = compare_to_benchmark({
+    "median_k_per_kind": {"q": 1.14, "k": 1.13, "v": 1.19, "o": 1.19}
+})
+print(cmp["nearest_neighbor"])  # nearest of the 12 benchmark entries
+```
+
+---
+
+## Dataset structure
+
+### 12 entries
+
+| `entry_id`    | Family   | Size  | Architecture  | QK-Norm | Tokens |
+|---------------|----------|-------|---------------|---------|--------|
+| pythia-70m    | Pythia   | 70m   | MHA-merged    | False   | 300B   |
+| pythia-160m   | Pythia   | 160m  | MHA-merged    | False   | 300B   |
+| pythia-410m   | Pythia   | 410m  | MHA-merged    | False   | 300B   |
+| pythia-1b     | Pythia   | 1B    | MHA-merged    | False   | 300B   |
+| pythia-6.9b   | Pythia   | 6.9B  | MHA-merged    | False   | 300B   |
+| olmo-1-7b     | OLMo-1   | 7B    | MHA-separate  | False   | 2.5T   |
+| olmo-2-7b     | OLMo-2   | 7B    | MHA-separate  | True    | 5T     |
+| llama-3-8b    | LLaMA-3  | 8B    | GQA-4:1       | False   | 15T    |
+| mistral-7b    | Mistral  | 7B    | GQA-4:1       | False   | 8T     |
+| qwen2.5-7b    | Qwen2.5  | 7B    | GQA-7:1       | False   | 18T    |
+| qwen2.5-14b   | Qwen2.5  | 14B   | GQA-5:1       | False   | 18T    |
+| qwen3-8b      | Qwen3    | 8B    | GQA-4:1       | True    | 36T    |
+
+### Columns (55 total)
+
+**Identifiers**: `entry_id`, `family`, `size`, `arch`
+
+**Architecture**: `n_q`, `n_kv` (head counts for GQA), `qk_norm`
+
+**Training hyperparameters**: `training_tokens`, `eta_peak` (peak learning rate), `lambda_wd` (weight decay), `T_steps`, `tau_iter = 1/(η·λ_wd)`, `T_over_tau` (Wang-Aitchison 2024 cycle ratio), `Physical_State`, `hp_confidence` (explicit / inferred / estimated)
+
+**Per-component Weibull fits** (median across blocks within each model):
+
+- Pythia entries (GeLU 2-projection FFN with merged W_qkv): populate `qkv`, `o`, `ffn_in`, `ffn_out`
+- Non-Pythia entries (SwiGLU 3-projection FFN with separate Q/K/V): populate `q`, `k`, `v`, `o`, `gate`, `up`, `down`
+
+Each component has four columns: `k_median_*`, `lambda_median_*`, `R2_median_*`, `R2_below_95_*` (count of blocks with R² < 0.95).
+
+`n_records` = total number of per-block Weibull fits aggregated.
+
+---
+
+## Key findings (paper §3)
+
+1. **Transmission Class** (FFN + W_o): the **median terminal k across components per entry**, then aggregated across the 12 entries, falls in **[1.186, 1.204]** with cross-family CV = **0.51%**. Shared across SwiGLU and GeLU activations, Pre-LN and QK-Norm placements, and model sizes 70M → 14B.
+
+2. **Selection Class** (W_q, W_k): k departs from initialization anchor (k_init ≈ 1.205), severity tracks attention storage architecture — **separately-stored MHA** (OLMo-1/2) deepest at k ∈ [0.76, 0.99], **GQA** (LLaMA-3, Mistral, Qwen2.5/3) milder at [1.10, 1.16], **Pythia merged W_qkv** transitional at [1.05, 1.18].
+
+3. **λ scaling**: terminal λ ∝ √(η/λ_wd) within Pythia, Pearson r = 0.94 across the three Transmission Class component kinds, directionally consistent with Fan et al. (2025) AdamW steady-state analysis.
+
+---
+
+## Files in this dataset
+
+| File | Purpose |
+|---|---|
+| `DATABASE_v9_1.csv` | Machine-readable dataset (55 columns × 12 entries) |
+| `DATABASE_v9_1.md` | Human-readable reference table |
+| `DATABASE_v9_1_report.md` | Per-entry sanity verification |
+| `populate_database_v9_1.py` | Internal regeneration script (dev-only; requires cascade v3 raw per-block fits not shipped here — see GitHub repo for the cascade pipeline) |
+| `LICENSE` | CC BY 4.0 license text |
+| `README.md` | This file |
+
+---
+
+## License
+
+[CC BY 4.0](LICENSE) — free for academic and commercial use with attribution.
+
+---
+
+## Citation
+
+```bibtex
+@article{ding2026twoparameterweibull,
+  title   = {A Two-Parameter Weibull Framework for Diagnosing Transformer Weight Distributions},
+  author  = {Ding, Tiexin},
+  journal = {arXiv preprint arXiv:2605.18898},
+  year    = {2026},
+  doi     = {10.48550/arXiv.2605.18898},
+  url     = {https://arxiv.org/abs/2605.18898}
+}
+```
+
+---
+
+**Author**: Tiexin Ding · NeuralCAE
+
+Issues / questions: [GitHub Issues](https://github.com/tiexinding/NPM-Weibull-public/issues)

populate_database_v9_1.py

@@ -0,0 +1,524 @@
+"""DATABASE_v9_1 populate — regenerate the 12-entry benchmark table.
+
+INTERNAL DEVELOPMENT SCRIPT. End users do NOT need to run this — the
+output files (DATABASE_v9_1.csv, .md, _report.md) are already committed
+alongside it. Use those files directly, or load via `npm-weibull-py`:
+
+    pip install npm-weibull-py
+    from npm_weibull import DATABASE_v9_1
+
+This script requires the cascade v3 raw per-block Weibull fit JSON
+files, which are NOT shipped in this repository (they total several GB
+of derived measurement data). It runs only on the author's development
+machine where the cascade pipeline output lives at the ROOT path
+hard-coded below.
+
+Output (written next to this script):
+  - DATABASE_v9_1.csv          (12 rows × 55 columns)
+  - DATABASE_v9_1.md           (human-readable reference table)
+  - DATABASE_v9_1_report.md    (per-entry sanity verification)
+
+Source data:
+  - cascade_v3_pull/data/derived/{model}_main_fit_per_component_v3.json
+  - cascade_v3_pull/data/derived/pythia-{size}-step143000_step143000_fit_per_component_v3.json
+  - derived/{llama-3-8b, mistral-7b, qwen2.5-7b, olmo2-7b-final}_main_fit_per_component_v3.json
+
+For the full cascade pipeline (raw checkpoints → per-block fits → this
+benchmark), see the project root README on GitHub.
+"""
+
+from __future__ import annotations
+
+import csv
+import json
+import statistics
+from pathlib import Path
+
+ROOT = Path(
+    "/home/dingdang-ws/wsl-projects/claudecode/NPM_v13_commit_260324/NPM_v13_complete/30_NPM_weibull/cascade_v2_20260502"
+)
+CV3 = ROOT / "cascade_v3_pull" / "data" / "derived"
+CV2_DERIVED = ROOT / "derived"
+# Output goes next to this script (release dir under NPM-Weibull-public/database_v9_1/).
+OUT = Path(__file__).resolve().parent
+OUT.mkdir(parents=True, exist_ok=True)
+
+
+# 12 family entry definitions + training hyperparameters for T/tau computation
+# Hyperparameters from each model's paper / HF config.json:
+#   eta_peak  : peak learning rate (after warmup)
+#   lambda_wd : weight decay coefficient
+#   T_steps   : total training steps (terminal)
+#   src_conf  : confidence (paper-explicit / inferred / estimated)
+#
+# Pythia: Biderman et al. 2023 (arXiv:2304.01373) Table 4
+# OLMo-1: Groeneveld et al. 2024 (arXiv:2402.00838) §3
+# OLMo-2: OLMo-2 paper (arXiv:2501.00656) §3
+# LLaMA-3: Meta Llama 3 paper (arXiv:2407.21783) §3.4
+# Mistral: tech report partial; estimates inferred from typical 7B recipe
+# Qwen2.5: Qwen2.5 paper (arXiv:2412.15115)
+# Qwen3:   Qwen3 paper (arXiv:2505.xxxxx)
+#
+# (entry_id, json_path, family, size, arch, n_q, n_kv, qk_norm, training_tokens,
+#  eta_peak, lambda_wd, T_steps, hp_confidence)
+ENTRIES = [
+    # ----- Pythia 5 size (terminal step143000) -- Biderman 2023 Table 4 -----
+    (
+        "pythia-70m",
+        CV3 / "pythia-70m-step143000_step143000_fit_per_component_v3.json",
+        "Pythia",
+        "70m",
+        "MHA-merged",
+        None,
+        None,
+        False,
+        "300B",
+        1.0e-3,
+        0.01,
+        143000,
+        "explicit",
+    ),
+    (
+        "pythia-160m",
+        CV3 / "pythia-160m-step143000_step143000_fit_per_component_v3.json",
+        "Pythia",
+        "160m",
+        "MHA-merged",
+        None,
+        None,
+        False,
+        "300B",
+        6.0e-4,
+        0.01,
+        143000,
+        "explicit",
+    ),
+    (
+        "pythia-410m",
+        CV3 / "pythia-410m-step143000_step143000_fit_per_component_v3.json",
+        "Pythia",
+        "410m",
+        "MHA-merged",
+        None,
+        None,
+        False,
+        "300B",
+        3.0e-4,
+        0.01,
+        143000,
+        "explicit",
+    ),
+    (
+        "pythia-1b",
+        CV3 / "pythia-1b-step143000_step143000_fit_per_component_v3.json",
+        "Pythia",
+        "1B",
+        "MHA-merged",
+        None,
+        None,
+        False,
+        "300B",
+        3.0e-4,
+        0.01,
+        143000,
+        "explicit",
+    ),
+    (
+        "pythia-6.9b",
+        CV3 / "pythia-6.9b-step143000_step143000_fit_per_component_v3.json",
+        "Pythia",
+        "6.9B",
+        "MHA-merged",
+        None,
+        None,
+        False,
+        "300B",
+        1.2e-4,
+        0.01,
+        143000,
+        "explicit",
+    ),
+    # ----- 6 cross-family -----
+    # OLMo-1 7B: Groeneveld 2024 §3.2 — peak LR 3e-4, wd 0.1, ~477k steps × global_batch 2160 × seq 2048 = ~2.1T tokens (paper says 2T+ but OLMo-1B used 3T tokens; 7B run is ~2.5T)
+    (
+        "olmo-1-7b",
+        CV3 / "olmo-7b-hf_main_fit_per_component_v3.json",
+        "OLMo-1",
+        "7B",
+        "MHA-separate",
+        32,
+        32,
+        False,
+        "2.5T",
+        3.0e-4,
+        0.1,
+        477000,
+        "explicit",
+    ),
+    # OLMo-2 7B: OLMo-2 paper — stage1 ~600k steps + stage2 annealing; here T_steps approx for stage1 saturation; LR 3e-4
+    (
+        "olmo-2-7b",
+        ROOT / "derived/olmo2-7b-final_main_fit_per_component_v3.json",
+        "OLMo-2",
+        "7B",
+        "MHA-separate",
+        32,
+        32,
+        True,
+        "5T",
+        3.0e-4,
+        0.1,
+        600000,
+        "inferred",
+    ),
+    # Llama-3 8B: Meta Llama 3 paper — peak LR ~3e-4, wd 0.1, 15T tokens, batch ~16M tokens → ~1M steps
+    (
+        "llama-3-8b",
+        ROOT / "derived/llama-3-8b_main_fit_per_component_v3.json",
+        "Llama-3",
+        "8B",
+        "GQA-4:1",
+        32,
+        8,
+        False,
+        "15T",
+        3.0e-4,
+        0.1,
+        1000000,
+        "inferred",
+    ),
+    # Mistral-7B: tech report partial — estimates inferred from typical Llama-style recipe
+    (
+        "mistral-7b",
+        ROOT / "derived/mistral-7b_main_fit_per_component_v3.json",
+        "Mistral",
+        "7B",
+        "GQA-4:1",
+        32,
+        8,
+        False,
+        "8T",
+        3.0e-4,
+        0.1,
+        500000,
+        "estimated",
+    ),
+    # Qwen2.5-7B: paper says peak LR 3e-4, wd 0.1, 18T tokens
+    (
+        "qwen2.5-7b",
+        ROOT / "derived/qwen2.5-7b_main_fit_per_component_v3.json",
+        "Qwen2.5",
+        "7B",
+        "GQA-7:1",
+        28,
+        4,
+        False,
+        "18T",
+        3.0e-4,
+        0.1,
+        1100000,
+        "inferred",
+    ),
+    # Qwen2.5-14B: GQA 40:8 = 5:1 (paper line 211 footnote: per-block extraction covers first 27 of 48 layers; aggregate-level entry); no QK-Norm (paper §QK-Norm figure caption)
+    (
+        "qwen2.5-14b",
+        CV3 / "qwen2.5-14b_main_fit_per_component_v3.json",
+        "Qwen2.5",
+        "14B",
+        "GQA-5:1",
+        40,
+        8,
+        False,
+        "18T",
+        3.0e-4,
+        0.1,
+        1100000,
+        "estimated",
+    ),
+    # Qwen3-8B-base: 36T tokens, similar recipe to Qwen2.5
+    (
+        "qwen3-8b",
+        CV3 / "qwen3-8b-base_main_fit_per_component_v3.json",
+        "Qwen3",
+        "8B",
+        "GQA-4:1",
+        32,
+        8,
+        True,
+        "36T",
+        3.0e-4,
+        0.1,
+        2200000,
+        "inferred",
+    ),
+]
+
+
+def compute_t_tau(eta_peak, lambda_wd, T_steps):
+    """Wang-Aitchison 2024 cycle ratio: T/tau_iter where tau_iter = 1/(eta * lambda_wd)."""
+    tau_iter = 1.0 / (eta_peak * lambda_wd)
+    return T_steps / tau_iter
+
+
+def classify_physical_state(t_over_tau):
+    """Physical state thresholds (Wang-Aitchison 2024 cycle ratio)."""
+    if t_over_tau >= 1.20:
+        return "Saturated"
+    if t_over_tau >= 0.80:
+        return "Near-saturated"
+    if t_over_tau >= 0.40:
+        return "Approaching"
+    if t_over_tau >= 0.25:
+        return "Partial"
+    return "Transition"
+
+
+# kinds for separate-Q/K models (SwiGLU/GeGLU FFN: gate/up/down)
+KINDS_SEPARATE = ["q", "k", "v", "o", "gate", "up", "down"]
+# kinds for Pythia (merged W_qkv + GeLU FFN: ffn_in/ffn_out, no gate)
+KINDS_MERGED = ["qkv", "o", "ffn_in", "ffn_out"]
+
+
+def median_of_kind(records, kind, field):
+    vals = [r[field] for r in records if r.get("kind") == kind]
+    return statistics.median(vals) if vals else None
+
+
+def count_low_R2(records, kind, threshold=0.95):
+    matching = [r for r in records if r.get("kind") == kind]
+    low = [r for r in matching if r.get("R2", 1.0) < threshold]
+    return len(low), len(matching)
+
+
+def load_entry(json_path):
+    if not json_path.exists():
+        return None
+    with open(json_path) as f:
+        d = json.load(f)
+    return d.get("per_component", [])
+
+
+def populate():
+    rows = []
+    sanity_lines = []
+
+    for (
+        eid,
+        jpath,
+        family,
+        size,
+        arch,
+        nq,
+        nkv,
+        qkn,
+        tok,
+        eta_peak,
+        lambda_wd,
+        T_steps,
+        hp_conf,
+    ) in ENTRIES:
+        records = load_entry(jpath)
+        if not records:
+            print(f"[WARN] missing data for {eid}: {jpath}")
+            continue
+
+        # T/tau and Physical State
+        t_over_tau = compute_t_tau(eta_peak, lambda_wd, T_steps)
+        phys_state = classify_physical_state(t_over_tau)
+
+        # Determine kinds based on architecture (Pythia merged W_qkv vs separate-Q/K)
+        is_merged = arch == "MHA-merged"
+        kinds = KINDS_MERGED if is_merged else KINDS_SEPARATE
+
+        row = {
+            "entry_id": eid,
+            "family": family,
+            "size": size,
+            "arch": arch,
+            "n_q": nq,
+            "n_kv": nkv,
+            "qk_norm": qkn,
+            "training_tokens": tok,
+            "eta_peak": eta_peak,
+            "lambda_wd": lambda_wd,
+            "T_steps": T_steps,
+            "tau_iter": round(1.0 / (eta_peak * lambda_wd)),
+            "T_over_tau": round(t_over_tau, 3),
+            "Physical_State": phys_state,
+            "hp_confidence": hp_conf,
+            "n_records": len(records),
+        }
+
+        for kind in kinds:
+            row[f"k_median_{kind}"] = median_of_kind(records, kind, "k")
+            row[f"lambda_median_{kind}"] = median_of_kind(records, kind, "lambda")
+            row[f"R2_median_{kind}"] = median_of_kind(records, kind, "R2")
+            low, tot = count_low_R2(records, kind)
+            row[f"R2_below_95_{kind}"] = f"{low}/{tot}" if tot else "n/a"
+
+        rows.append(row)
+
+        # Sanity logging
+        sanity_lines.append(f"### {eid} ({family} {size}, {arch})")
+        sanity_lines.append(f"  records: {len(records)}; tokens: {tok}; QK-Norm: {qkn}")
+        if is_merged:
+            sanity_lines.append(
+                f"  k_median(qkv) = {row['k_median_qkv']:.4f}, R2(qkv) = {row['R2_median_qkv']:.4f}, low-R2: {row['R2_below_95_qkv']}"
+            )
+        else:
+            sanity_lines.append(
+                f"  k_median(q)  = {row['k_median_q']:.4f}, R2(q)  = {row['R2_median_q']:.4f}, low-R2: {row['R2_below_95_q']}"
+            )
+            sanity_lines.append(
+                f"  k_median(k)  = {row['k_median_k']:.4f}, R2(k)  = {row['R2_median_k']:.4f}, low-R2: {row['R2_below_95_k']}"
+            )
+            sanity_lines.append(
+                f"  k_median(v)  = {row['k_median_v']:.4f}, R2(v)  = {row['R2_median_v']:.4f}"
+            )
+            sanity_lines.append(
+                f"  k_median(o)  = {row['k_median_o']:.4f}, R2(o)  = {row['R2_median_o']:.4f}"
+            )
+
+        def fnum(v):
+            return f"{v:.4f}" if isinstance(v, (int, float)) else "n/a"
+
+        sanity_lines.append(
+            f"  k_median(gate)={fnum(row.get('k_median_gate'))}, k_median(up)={fnum(row.get('k_median_up'))}, k_median(down)={fnum(row.get('k_median_down'))}"
+        )
+        sanity_lines.append("")
+
+    # ===== Write CSV =====
+    csv_path = OUT / "DATABASE_v9_1.csv"
+    if rows:
+        # Union of all keys (separate + merged)
+        keys = list(rows[0].keys())
+        for r in rows[1:]:
+            for k in r:
+                if k not in keys:
+                    keys.append(k)
+        with open(csv_path, "w", newline="") as f:
+            writer = csv.DictWriter(f, fieldnames=keys, extrasaction="ignore")
+            writer.writeheader()
+            for r in rows:
+                writer.writerow({k: r.get(k, "") for k in keys})
+        print(f"[saved CSV] {csv_path}  ({len(rows)} rows)")
+
+    # ===== Write markdown reference table =====
+    md_path = OUT / "DATABASE_v9_1.md"
+    with open(md_path, "w") as f:
+        f.write("# DATABASE_v9_1 — Reference Table (12 family entries)\n\n")
+        f.write("**Source**: cascade v3 (per-component Weibull fit, mid-80% trim)\n")
+        f.write("**Generated**: by populate_database_v9_1.py\n\n")
+        f.write("## 12 entries (5 Pythia size + 7 cross-family) — k median per component\n\n")
+        f.write(
+            "| Entry | Family | Size | Architecture | n_q/n_kv | QK-Norm | tokens |"
+            " k_q | k_k | k_v | k_o | k_gate | k_up | k_down | low-R²(q) |\n"
+        )
+        f.write("|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n")
+        for r in rows:
+            qk = "yes" if r["qk_norm"] else "no"
+            nqkv = f"{r['n_q']}/{r['n_kv']}" if r["n_q"] else "merged"
+            kq = r.get("k_median_q") or r.get("k_median_qkv") or 0
+            kk = r.get("k_median_k") or "—"
+            kv = r.get("k_median_v") or "—"
+            ko = r.get("k_median_o") or "—"
+            kg = r.get("k_median_gate") or 0
+            ku = r.get("k_median_up") or 0
+            kd = r.get("k_median_down") or 0
+            lowq = r.get("R2_below_95_q", r.get("R2_below_95_qkv", "—"))
+
+            def fmt(v):
+                return f"{v:.4f}" if isinstance(v, (int, float)) else str(v)
+
+            f.write(
+                f"| {r['entry_id']} | {r['family']} | {r['size']} | {r['arch']} | {nqkv} | {qk} | {r['training_tokens']} | "
+                f"{fmt(kq)} | {fmt(kk)} | {fmt(kv)} | {fmt(ko)} | {fmt(kg)} | {fmt(ku)} | {fmt(kd)} | {lowq} |\n"
+            )
+        f.write("\n")
+
+        # ===== T/tau + Physical State table =====
+        f.write(
+            "## Training hyperparameters + T/τ + Physical State (Wang-Aitchison 2024 cycle ratio)\n\n"
+        )
+        f.write(
+            "τ_iter = 1/(η · λ_wd) — EMA iteration time-constant.  T/τ_iter = T_steps / τ_iter — completed EMA cycles.\n\n"
+        )
+        f.write(
+            "| Entry | η_peak | λ_wd | T_steps | τ_iter | **T/τ** | **Physical State** | hp source |\n"
+        )
+        f.write("|---|---|---|---|---|---|---|---|\n")
+        for r in rows:
+            f.write(
+                f"| {r['entry_id']} | {r['eta_peak']:.1e} | {r['lambda_wd']} | {r['T_steps']} | "
+                f"{int(r['tau_iter'])} | **{r['T_over_tau']:.2f}** | **{r['Physical_State']}** | {r['hp_confidence']} |\n"
+            )
+        f.write("\n**Physical State thresholds** (Wang-Aitchison 2024 cycle ratio):\n\n")
+        f.write("- Saturated: T/τ ≥ 1.20\n")
+        f.write("- Near-saturated: 0.80 ≤ T/τ < 1.20\n")
+        f.write("- Approaching: 0.40 ≤ T/τ < 0.80\n")
+        f.write("- Partial: 0.25 ≤ T/τ < 0.40\n")
+        f.write("- Transition: T/τ < 0.25\n\n")
+        f.write("**hp source confidence**:\n\n")
+        f.write("- *explicit*: paper Table / official tech report directly states the value\n")
+        f.write("- *inferred*: paper §3 states a quantity from which we derive it (e.g. tokens × batch / seq → steps)\n")
+        f.write("- *estimated*: paper does not publish; same-family typical recipe used as fallback\n\n")
+
+        # ===== Public verification section =====
+        f.write("---\n\n## Verification\n\n")
+        f.write(
+            "Per-entry per-component sanity check is recorded in "
+            "[`DATABASE_v9_1_report.md`](DATABASE_v9_1_report.md). "
+            "All entries pass R² ≥ 0.99 on the Transmission Class components.\n\n"
+        )
+
+        # Compute aggregated Transmission band (median across components per
+        # entry, then aggregated across entries) — public-facing summary.
+        def fmt_or_dash(v):
+            return f"{v:.4f}" if isinstance(v, (int, float)) else "—"
+
+        per_entry_med = []
+        for r in rows:
+            # Transmission Class = FFN + W_o per paper §3
+            comps = []
+            if r.get("k_median_qkv"):  # Pythia merged: ffn_in + ffn_out + o
+                for kind in ("ffn_in", "ffn_out", "o"):
+                    v = r.get(f"k_median_{kind}")
+                    if v is not None:
+                        comps.append(v)
+            else:  # SwiGLU: gate + up + down + o
+                for kind in ("gate", "up", "down", "o"):
+                    v = r.get(f"k_median_{kind}")
+                    if v is not None:
+                        comps.append(v)
+            if comps:
+                per_entry_med.append(statistics.median(comps))
+
+        if per_entry_med:
+            mean_k = statistics.mean(per_entry_med)
+            cv_pct = statistics.stdev(per_entry_med) / mean_k * 100
+            f.write(
+                f"**Transmission Class aggregated band** (median across "
+                f"components per entry, then aggregated across the "
+                f"{len(per_entry_med)} entries): "
+                f"k ∈ [{min(per_entry_med):.4f}, {max(per_entry_med):.4f}], "
+                f"cross-family CV = {cv_pct:.2f}%.\n\n"
+            )
+            f.write(
+                "See paper §3 for the strict-band definition and trim "
+                "protocol. For per-block raw fits and the cascade "
+                "pipeline that produces this table, see the "
+                "`npm-weibull-py` repository on GitHub.\n"
+            )
+
+    print(f"[saved MD] {md_path}")
+
+    # ===== Sanity report =====
+    rpt = OUT / "DATABASE_v9_1_report.md"
+    with open(rpt, "w") as f:
+        f.write("# DATABASE_v9_1 sanity verify report\n\n")
+        f.write("\n".join(sanity_lines))
+    print(f"[saved report] {rpt}")
+
+
+if __name__ == "__main__":
+    populate()