AtlasOpenCoder / validation /paper2 /A38_POPULATION.json
piotrwilam's picture
AtlasOpenCoder v1: the battery, the artefacts and both papers' manifests
9b7bd53
Raw History Blame Contribute Delete
1.02 kB
{
"what": "the population behind A38's Spearman -0.813, recorded so the figure the paper quotes in \u00a74.6 has an artefact rather than only prose",
"why": "CC read '1,404 pairs' as reference pairs and was wrong; the number appears only in prose in experiments/step6_token.py:11 and experiments/step8_graph.py:12",
"factor_band": [
1.25,
2.0
],
"n_pretraining_references": 35,
"n_point_partner_pairs_inside_the_band": 117,
"n_scope_varying_constructs": 12,
"product": 1404,
"matches_1404": true,
"not_reference_pairs": {
"C(35,2)": 595,
"C(61,2)": 1830,
"note": "1,404 is not triangular; the nearest are 1,378 (53) and 1,431 (54)"
},
"provenance": {
"inputs": [
"validation/carriers/*.npz for the reference list",
"experiments/construct_token_loss.py FACTOR_BAND"
],
"reference": null,
"battery_prompt_hash": "570b42db13f8fe5772b8d1a9ad7d55d6fa1279dae5948ad94e8363789bcbc1ad",
"code_commit": "3325b25",
"design": "9-Paper-1/2610-18Oct_PAPER-1-Design.md sha256 47f6a664\u2026"
}
}