aimflow/EVAL.md

aimflow (v1) — evaluation

QwenPI + hybrid flow-aim head. Weight: steps_5000.pt. ⚠️ v1's eval aim-cosine is a normalization artifact (see ../README.md), so expect unreliable live aim. Needs Qwen3-VL for the model server. Full guide: ../EVALUATION.md.

Run (from inside this folder; game server + client already up):

mkdir -p run/checkpoints && cp config.yaml dataset_statistics.json run/ && cp steps_5000.pt run/checkpoints/
python -m gamevla.envs.eval.run_eval \
  --env cs2/5e_mirage_prefire/connector_to_a_site --endpoint <host:port> \
  --policy examples.CS2.aimflow.evaluation.policy:build_policy \
  --policy-kwargs '{"ckpt":"run/checkpoints/steps_5000.pt","port":10093,"sensitivity":1.0}' \
  --episodes 3 --bot-count 5

Defaults (policy.py): steps_per_bin=1, history_len=2. Mouse auto-derived from sensitivity.