Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

2.8 KB
Raw Blame

Source and evidence appendix

This appendix contains small, auditable inputs needed to understand or reproduce the two vision-adapter experiments. It intentionally excludes model weights, optimizer tensors, dataset images, API tokens, SSH material, caches, and virtual environments.

Layout

  • deepseek-v4-flash/: a curated snapshot of the local DeepSeek implementation, including configs, training/inference code, SGLang extension code, scripts, tests, and design notes.
  • laguna-xs-2.1/private-training-archive/: the Laguna private Hugging Face archive's README, training log, final run summary/state, archive SHA list, and the two code snapshots stored with the run.
  • published-artifacts/: copies of the public model cards and machine-readable adapter manifests at the revisions listed in ARTIFACTS.json.
  • SOURCE_SNAPSHOT.json: provenance and caveats for the copied local source.
  • ARTIFACTS.json: compact final repository/component manifest.
  • SHA256SUMS: hashes for every file in this appendix except SHA256SUMS itself.

Source snapshot caveat

The local DeepSeek project was not a clean Git checkout when this appendix was created. Its base commit was 285a35a1c5cbff3251418013a0671b25ef78488a, with modified training/guard files and uncommitted serving, inference, SGLang, and test additions. The snapshot is therefore described as base commit plus working-tree changes, not as the contents of that commit. The exact dirty-file inventory is in SOURCE_SNAPSHOT.json.

The five packaged SGLang runtime/documentation files and three focused regression tests listed in SOURCE_SNAPSHOT.json were later synchronized byte-for-byte from the final public DeepSeek NVFP4 revision. That bounded sync does not turn the rest of the historical snapshot into a clean public-package checkout.

This distinction matters because the local README contains a historical sentence saying the 100K run had not started, while the later public model cards and final projectors show that it subsequently completed. For final artifact status, prefer ARTIFACTS.json and the commit-pinned Hugging Face pages.

Laguna archive caveat

The copied Laguna run-summary.json describes the resumed process only. Training had first reached step 384, then restarted from the last durable step-380 checkpoint. Consequently, the final summary's examples field is cumulative, while its elapsed_seconds, examples_per_hour, and first/minimum loss fields cover only the resumed segment. See the runbook for the correct interpretation.

Verify the appendix

From this directory:

shasum -a 256 -c SHA256SUMS

To verify a large Hugging Face artifact, download the exact revision into external storage and compare its SHA-256 with ARTIFACTS.json. Do not place multi-gigabyte weights in this documentation tree.