Juergen

Juergen contains two SFT data streams and one evaluation suite:

• the canonical Crowd-Cast pipeline, annotated with describe_extract and trained with deltatype_v2;
• the CUA-Gym action-format parity pipeline, trained with the normalized relative ordered_events_v3 grammar;
• the standalone CUA micro-eval under eval/.

Training is owned by Omegalax. Desktop/QEMU lifecycle and Slurm deployment are separate repositories.

Run the whole suite with:

uv run --locked --extra dev pytest -q

Run the micro-eval tests and evaluator from its independent lock:

uv run --project eval --locked pytest -q eval
uv run --project eval --locked python eval/cua_micro_eval.py \
  --model_path <hf-checkout> --output_dir <dir> --qemu_image <qcow2>

See pipeline/README.md and eval/README.md for the artifact contracts and runtime options.