Skip to content

Project History

Non-release project history lives here. Release and incremental release notes live in docs/versions/<semver>.md.

2026-05-24 — supervised notebook heartbeat

Taught repo-root train.py to emit stage start/finish lines and a 60-second keepalive while uv run jupyter nbconvert --execute ... is still active. This makes detached training-manager runs less opaque and gives the manager coarse progress events even when the notebook itself is quiet.

2026-05-24 — training-manager entrypoint

Added repo-root train.py as a stable supervised entrypoint for the full notebook pipeline. The wrapper preserves the repo's uv-only rule by running uv sync and then executing full_llm_pipeline.ipynb via uv run jupyter nbconvert --execute. This makes the repo manageable by external supervisors such as training-manager without pretending the notebook was already a clean standalone CLI.

2026-04-24 — Checklist no longer stores round logs

Moved append-only history out of docs/LLM_CHECK.md. The checklist is now checklist-only and explicitly says not to add round logs there. Incremental release notes live in their own files such as docs/versions/0.1.1.md, docs/versions/0.1.2.md, and docs/versions/0.2.1.md; non-release project history lives in this file.

2026-04-24 — Canonical long-running rules

Fast-forwarded the AGENTS.md submodule from 6cfe3bd to 6b6c218, then added and pushed canonical rule commit 51b8f27 (docs: require progress for long-running work). The rule requires visible progress for work that may exceed 1 minute, warns that 5 minutes of silence may be terminated, and requires durable checkpoints for work that may exceed 1 hour. Lavender-2 records the updated submodule pointer separately from project code.

2026-04-22 — uv-only local environment

Bumped local development from Anaconda to uv as the single source of truth. Disabled auto_activate_base in the user .condarc, ran uv sync, and registered the Lavender (uv .venv) Jupyter kernel. The rule is codified in docs/AGENTS.md; the owner task tracker is docs/TODO.md.

2026-04-22 — Wikipedia payload cap

Added a 32 GiB byte cap on the Wikipedia training payload in download_data.py. The script keeps wikimedia/wikipedia on the memory-mapped non-streaming path, scans text bytes, and keeps only the fitting prefix before save_to_disk.

2026-04-14 — Training docs restructure

Restructured docs around a Training section with a v0/ subfolder. Created component pages for data, tokenizer, architecture, MoE, training loop, checkpointing, and inference; rewrote docs/index.md; updated the VitePress sidebar and nav.

2026-04-14 — Cloudflare Workers Builds docs deploy

Cloudflare deploys used Workers Builds (npx wrangler deploy) rather than classic Pages. Added docs/wrangler.jsonc for an assets-only Worker and rewrote the Cloudflare deployment notes around that flow.

2026-04-14 — Cloudflare hosting switch

Swapped docs hosting from Vercel to Cloudflare after a Vercel account constraint. Removed Vercel config, added Node 22 metadata, and updated deployment references.

2026-04-14 — VitePress docs pipeline

Retired the mkdocs + GitHub Pages workflow and converted docs/ into a scoped VitePress project. Docs tooling remains under docs/.

2026-04-13 — Canonical agent submodule

Added github.com/Trance-0/AGENTS.md as the AGENTS.md/ submodule, created the initial Lavender-2 checklist and CLAUDE.md wrapper, and later fast-forwarded the submodule to inherit canonical docs/TODO and commit-hygiene rules.

2026-04-13 — Initial docs tree

Built the initial docs/ tree, pinned the v0.1.0 baseline, added deployment / development notes, and tightened .gitignore for logs and common caches.

2026-04-13 — Wikipedia OOM fix

Fixed unlimited Wikipedia download OOMs by switching wikimedia/wikipedia to streaming=False and saving non-streaming datasets via memory-mapped Arrow rather than accumulating all rows in RAM. Verified end-to-end with --limit 100.