Skip to content

MACD 3.1: working rewrite, switchable backends, real LoRA training - #1

Merged
bobathefetts merged 15 commits into
mainfrom
v3.1-rewrite
Oct 6, 2026
Merged

bobathefetts merged 15 commits into
mainfrom
v3.1-rewrite

Conversation

@bobathefetts

Copy link
Copy Markdown
Owner

Replaces the 3.0.0 code, which failed at import under its own pinned dependencies, with a rewrite that runs and is tested.

What changes: the project moves from the macd_v3/ subfolder to the repository root. The evolution loop now keeps and scores its offspring. Train, dev and test data are separate and checked for overlap. Backends are switchable in YAML (mock, any OpenAI-compatible server, in-process Hugging Face with PEFT LoRA). A distilled adapter is kept only if the dev score does not regress. Runs are reproducible from a seed.

Docs: the README now documents only what the code does, with a quickstart that needs no GPU, a status section listing what is and is not verified, and a roadmap. Development status notes, compiled .pyc files and local editor settings are removed. See CHANGELOG.md for the full list.

Testing: 51 tests pass locally on Python 3.10 to 3.13 with ruff clean. CI runs lint and tests on four Python versions, plus a job that performs real LoRA training on CPU against a tiny Llama built in the test.

Not yet exercised: a full-size model on a GPU, 4-bit loading, and vLLM runtime adapter loading.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XUMgK7AFz2EkWWU99WeTkM

@bobathefetts
bobathefetts merged commit 8f04c67 into main Oct 6, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant