alphat01/scaling_law

Existing artifacts for the breadth–depth composition study

This curated release contains existing training code, the two processed robot task stores, and the original logs/configurations underlying the paper's main grid and clean E=24 sensitivity runs. Access requires a request and manual author approval. No policy checkpoints are included.

Download and extract

All selected code, task data, experiment logs, derived tables and audit records are delivered in one file: scaling_law.zip. Extract it with a ZIP64-capable tool, for example unzip scaling_law.zip. The directories below are inside that ZIP. Existing task ZIP stores are preserved byte-for-byte inside it. No policy checkpoints are included. The repository remains public with manual access approval.

Contents inside scaling_law.zip

• code/: the existing Data-Scaling-Laws training framework and scripts; runtime caches and old W&B runs are excluded.
• data/dataset/{arrange_mouse,pour_water}/: original dataset.zarr.zip stores and count.txt source-count files.
• results/main/: 96 completed runs, 80 task–E–d configurations, eight three-seed configurations; validation sources 17–32.
• results/sensitivity_E24/: 24 clean E=24 runs; validation sources 25–32. These must not be pooled with the strict main evaluation without noting the different boundary.
• results/derived/: newly extracted final-epoch metrics and configuration means, with relative paths back to their raw logs. These are reconstruction/audit outputs, not a substitute for the missing author-frozen analysis archive.
• audit/: selection rationale, excluded material and an exact source-file manifest.
• environment/: existing archived environment records.

The processed robot data and training-framework foundation come from Lin et al., Data Scaling Laws in Imitation Learning for Robotic Manipulation (ICLR 2025). Existing source attribution and licensing files are retained. This release does not assign a new license to upstream material.

Important distinctions

The original E=24 runs with validation sources 17–32 overlap their training sources and are excluded. The empty-metric legacy CSV exports, duplicate archives, caches, trash, checkpoints, and MOCK LIBERO pilot rollouts are excluded. Publication-level prediction, figure-generation and published-score reanalysis scripts named in the supplement were not located; this is a release of the existing training/data/run artifacts, not a claim that every publication analysis is reproduced.

Paths

The archived Hydra configs preserve their original absolute paths for provenance. When using the code from this repository, point task.dataset_path at the corresponding ../data/dataset/<task>/dataset.zarr.zip store. Original directory symlinks are not followed into duplicated data copies. The existing ZIP files are the original Zarr stores used by training; no new data archives are created for upload.

Verified archive checksum

SHA-256 (scaling_law.zip): 8a89809cb7642c9dfb66895051750fb6214b4753bb4f0e8048fc6dbc0f13df35

Remote size and SHA-256 verified at 2026-11-01T07:35:52Z.