blitz-cli 0.3.0

Developer CLI for Blitz: pull a workflow's training data + base-model recommendation and scaffold a runnable QLoRA training container.

blitz-cli

Developer CLI for Blitz. Pull a workflow's captured traces as a fine-tuning dataset + the recommended open base model, and scaffold a self-contained, runnable QLoRA training container — then close the loop by grading the trained student against the held-out eval set.

pip install blitz-cli

export BLITZ_API_KEY=blz_...   # a read-scoped project key (mint one in the dashboard)
blitz scaffold -p proj_abc -w mechanic-assistant -o ./train

cd train
make build && make train && make eval

blitz scaffold writes ./train with data/dataset.jsonl, data/evalset.jsonl, config.json (derived from the base-model recommendation), and a Dockerfile + train.py + eval.py + Makefile. Bring your own NVIDIA GPU.

This package is intentionally dependency-free (stdlib + the Blitz HTTP API). The heavy ML stack (torch / transformers / trl / peft) is pinned only in the generated project's requirements.txt, installed inside the training image.

The complementary tracing SDK (pip install blitz-sdk, import blitz) lives in ../blitz-sdk.