Datasets:
DOI:
License:
File size: 1,023 Bytes
9b7bd53 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 | {
"what": "whether 'linear in the learning rate' and 'linear in log-steps' are separable on the planned 1,000-step Pythia grid, from the schedule alone",
"schedule": {
"total_steps": 143000,
"lr_max": 0.001,
"lr_min_fraction": 0.1,
"warmup_steps": 1430,
"shape": "linear warmup then cosine decay"
},
"grid": {
"n_points": 143,
"first": 1000,
"last": 143000,
"spacing": 1000
},
"pearson_r_lr_vs_log_steps": -0.8356681943232172,
"variance_inflation_factor": 3.3150050132037077,
"residual_sd_of_lr_after_log_steps": 0.5492346210833527,
"separable": true,
"reading": "the two predictors are separable enough on this grid for a fit to attribute an effect",
"minutes": 9.381771087646485e-06,
"provenance": {
"inputs": [
"Pythia's published schedule; no model, no run"
],
"reference": null,
"battery_prompt_hash": "570b42db13f8fe5772b8d1a9ad7d55d6fa1279dae5948ad94e8363789bcbc1ad",
"code_commit": "1d8310f",
"design": "9-Paper-1/2610-18Oct_PAPER-1-Design.md sha256 47f6a664\u2026"
}
} |