taf-agent / data /e4_gamma /allenai--OLMo-7B_random.json
karlexmarin's picture
feat: ship paper artefacts + CLI diagnostic alongside browser tool
535348a
raw
history blame
3.05 kB
{
"model": "allenai/OLMo-7B",
"theta": 10000,
"gamma_pred": 0.99999561666838,
"C_theory": 9.2103,
"distances": [
10,
20,
30,
50,
100,
200,
500,
1000,
1950
],
"distances_fit": [
30,
50,
100,
200,
500,
1000,
1950
],
"d_min_fit": 30,
"n_prompts_per_distance": 150,
"seeds": [
42,
123,
7
],
"attn_by_distance": {
"10": {
"mean_across_seeds": 0.011078201784855789,
"std_across_seeds": 0.0004568002389428613,
"per_seed_means": [
0.01044130727648735,
0.011490317210555077,
0.011302980867524941
]
},
"20": {
"mean_across_seeds": 0.008723362791869376,
"std_across_seeds": 0.0009938605218391735,
"per_seed_means": [
0.00813545612928768,
0.007911688508465886,
0.010122943737854561
]
},
"30": {
"mean_across_seeds": 0.007300337539571854,
"std_across_seeds": 0.00046900116814976243,
"per_seed_means": [
0.006724534671132763,
0.007303139980261525,
0.007873337967321276
]
},
"50": {
"mean_across_seeds": 0.005650945112316144,
"std_across_seeds": 0.00027488861469437606,
"per_seed_means": [
0.006032803847144047,
0.005396900329117974,
0.0055231311606864135
]
},
"100": {
"mean_across_seeds": 0.004058174822582967,
"std_across_seeds": 0.0005285617905847457,
"per_seed_means": [
0.0040144564071670176,
0.004726279242895544,
0.0034337888176863393
]
},
"200": {
"mean_across_seeds": 0.001906132831548651,
"std_across_seeds": 0.000362056287267553,
"per_seed_means": [
0.0016092082873607676,
0.0016933417382339637,
0.002415848469051222
]
},
"500": {
"mean_across_seeds": 0.0015021604990276196,
"std_across_seeds": 0.000235915454698382,
"per_seed_means": [
0.0014356055110692977,
0.0012523090862669051,
0.0018185668997466565
]
},
"1000": {
"mean_across_seeds": 0.0013560072134714575,
"std_across_seeds": 0.0008188443016643402,
"per_seed_means": [
0.0003828349825926125,
0.0012990302301477642,
0.0023861564276739956
]
},
"1950": {
"mean_across_seeds": 0.0006863480689935387,
"std_across_seeds": 0.00031650373924358707,
"per_seed_means": [
0.0009987773909233512,
0.0002525494305882603,
0.0008077173854690045
]
}
},
"fit_power_law": {
"gamma": 0.5411598042384096,
"log_A": -3.1051096313732693,
"R2": 0.96368,
"n_points": 7
},
"fit_exponential": {
"lambda": 0.0010459025390052495,
"log_A": -5.463978356707257,
"R2": 0.746322,
"n_points": 7
},
"delta_R2_power_minus_exp": 0.2174,
"gamma_CI_95": {
"lo": 0.448109235206576,
"hi": 0.6342278689106579
},
"decision": "UNCLEAR: \u03b3=0.541 outside all expected ranges"
}