Modelpublic

JamesK2W/gamevla

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes

Model Card

gamevla — CS2 VLA / behavioral-cloning checkpoints

Checkpoints for real-environment testing. One folder per experiment; each folder has config.yaml, dataset_statistics.json, the selected checkpoint(s), and an EVAL.md with the exact run command. For experiments with more than one saved checkpoint we keep the final step and one intermediate step; single-checkpoint experiments keep the only one.

▶ To run these in the CS2 environment, see [EVALUATION.md](EVALUATION.md) (score = mean total episode reward) and each folder's EVAL.md.

aim_cos = eval mouse-aim cosine similarity (the aim quality metric; ~0 means aim did not learn). key_f1 = eval h0 key macro-F1. attack_f1 = eval firing F1.

folder	checkpoint	step	aim_cos	key_f1	attack_f1
`csbc`	`steps_5000.pt` (best eval)	5000	0.00	0.285	0.156
`csbc`	`steps_15000.pt` (final)	15000	0.00	0.265	0.179
`nitrogen_starvla`	`steps_30000.pt` (mid)	30000	0.00	0.179	0.058
`nitrogen_starvla`	`final_model.pt` (final, complete 60k)	60000	0.01	0.191	0.065
`aimflow`	`steps_5000.pt` (only)	5000	0.475 ⚠️	0.143	0.042
`aimflow_v2`	`steps_10000.pt` (mid)	10000	-0.01	0.151	0.027
`aimflow_v2`	`steps_15000.pt` (final)	15000	0.01	0.158	0.021
`aimflow_v3`	`steps_30000.pt` (mid)	30000	-0.03	0.182	0.117
`aimflow_v3`	`steps_55000.pt` (final)	55000	0.04	0.157	0.025

Notes:

—⚠️ `aimflow` (v1) aim_cos ≈ 0.475 is a metric artifact, not real aim: v1 used a different mouse normalization, so its aim_cos is NOT comparable to v2/v3 (which read ~0). Across all experiments here, aim is not convincingly learned — treat it as an open question to verify in the real environment.
—csbc peaks early (step 5000) then overfits on keys; steps_15000 is more trained but slightly worse.
—nitrogen_starvla/final_model.pt is the only fully-completed (60k-step) run.