zsqzz/llm-memory-editability

LLM Memory: Learning and Editability — Experiment Archive

Data, predictions, numerical arrays, logs, research reports, frozen execution sources, and audits for zhusq20/llm-memory-editability. GitHub stores current source code, tests, configuration, and development instructions.

Layout

• data/: experiment datasets, preprocessing products, and tokenizers.
• results/: raw predictions, learning curves, arrays, logs, and run status.
• docs/: research reports, protocols, frozen sources, figures, and audits.
• Research index: the detailed project progress overview.
• ARCHIVE_MANIFEST.json: exact paths, byte sizes, SHA256 hashes, and excluded-file inventory.

Original project-relative paths are preserved. Use scripts/download_experiment_artifacts.py from the GitHub repository to restore data, results, and documentation locally.

Model weights, optimizer states, Docker image archives, Python environments, and internal W&B SDK caches are excluded. The user's 200 MB per-file limit applies only to GitHub; larger data files are included here.

This is a file-level archive. Runs that were active during capture may have newer local outputs; the archive is not an atomic model checkpoint and does not imply completion. Consult each batch's status, frozen protocol, data-source manifests, and interpretation boundaries. Upstream data retains its original licenses and attribution requirements.

Numerical array bundles

NumPy arrays are stored in per-experiment array-archives/*.tar.gz bundles to avoid tens of thousands of tiny LFS objects. ARRAY_ARCHIVES.json maps each original path to its bundle. ARCHIVE_MANIFEST.json retains every original file's SHA256 and byte size, plus the physical upload inventory. The GitHub download script selects the required bundles, verifies them, and restores selected arrays to their original paths. All array members were independently read back and hash-checked before upload.

Snapshot update — 2026-10-06

Incremental file capture: 2026-10-06T03:47:27.588435+00:00 through 2026-10-06T03:55:35.200178+00:00. Source code revision: f298768b6843.

New and updated experiment data, predictions, logs, reports, and frozen execution materials are included. See current research results and each archived controller state for the captured progress. Running experiments may produce later files. This snapshot does not imply that active runs have finished. Earlier artifacts are retained when no replacement is captured.

Snapshot update — 2026-10-08

Incremental file capture: 2026-10-08T05:10:32.154282+00:00 through 2026-10-08T05:39:01.781585+00:00. Source code revision: eb6fa2872dae.

New and updated experiment data, predictions, logs, reports, and frozen execution materials are included. See current research results and each archived controller state for the captured progress. Running experiments may produce later files. This snapshot does not imply that active runs have finished. Earlier artifacts are retained when no replacement is captured.

Credential-like source files are omitted and recorded in excluded_files in the archive manifest; the original local files are preserved. Numerical arrays are hash-checked and bundled with their original paths.

Live worker log snapshots

Two live worker logs are captured as exact, double-hash-verified prefixes through 2026-10-08T05:56:59.962675+00:00. Each prefix stops at the source file size observed at open; subsequent output is outside this snapshot. Capture times and SHA256 are recorded in the manifest.