Datasets:
DOI:
License:
| { | |
| "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" | |
| } | |
| } |