hab-swe/20261019-swebv-loopback-medcli-v2-grpo

SWE-bench Verified, gateway-only agent egress: MedCLI-V2 GRPO checkpoints

Full SWE-bench Verified (Harbor Hub swe-bench/swe-bench-verified revision 2, 500 tasks x 1 attempt) with --agent-egress loopback: while the agent works, its sandbox reaches only the model gateway (the verifier keeps the network). The score is resolved instances / 500 under the pinned Harbor adapter's verifier (not a leaderboard submission), withheld unless all 500 trials have valid verdicts. Run through Polar's Harbor runner (examples/runs/20261019-tb21-swebv-medcli-v2-grpo in ProRL-Agent-Server-v2, run id 20261019-swebv-loopback-medcli-v2-grpo). Agent: vanillux, 1024 steps, 32768 maximum completion tokens per call, preserved thinking, vision for the vision-language checkpoints; requests at temperature 1.0, top-p 0.95, top-k 20 with the chat-template switches enable_thinking, reasoning_effort xhigh. SGLang 0.5.15.post1 on eight B200s, TP2 x DP4 at context 262144; concurrency 32 with a 64-CPU sandbox budget; original task timeouts. Unscored (errored) trials get one retry pass; a job still incomplete after it is uploaded as it is.

Timeouts as 0 counts every trial whose agent ran out of its time budget as failed (Harbor still verifies them, and some pass). Unscored trials count as 0 in both columns.

Model	Checkpoint	Score	Timeouts as 0	Successful / valid trials	Timeouts (passed)	Serving	Tool-call parser	Vision	Results
medcli-v2-grpo-27b-from35-i77	20261005-medcli-v2-grpo-qwen3.8-27b-2node-ctm-top50-ehr-prec20-from35-iter_0000077-complete	0.7520	0.750	376/500	4 (1)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-27b-from35-i77/
medcli-v2-grpo-27b-scratch-i41	20261007-medcli-v2-grpo-qwen3.8-27b-2node-ctm-top50-ehr-prec20-scratch-iter_0000041-complete	withheld: 1/500 unscored	0.764	388/499	13 (6)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-27b-scratch-i41/
medcli-v2-grpo-27b-scratch-i83	20261007-medcli-v2-grpo-qwen3.8-27b-2node-ctm-top50-ehr-prec20-scratch-iter_0000083-complete	withheld: 2/500 unscored	0.744	375/498	8 (3)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-27b-scratch-i83/
medcli-v2-grpo-9b-scratch-i41	20261012-medcli-v2-grpo-qwen3.5-9b-q38-ctm-top50-ehr-prec10-scratch-iter_0000041-complete	withheld: 1/500 unscored	0.576	290/499	9 (2)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-9b-scratch-i41/
medcli-v2-grpo-9b-scratch-i83	20261012-medcli-v2-grpo-qwen3.5-9b-q38-ctm-top50-ehr-prec10-scratch-iter_0000083-complete	withheld: 1/500 unscored	0.572	286/499	11 (0)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-9b-scratch-i83/
medcli-v2-grpo-9b-0918-i79	20260918-medcli-v2-grpo-qwen3.5-9b-q38-hf-no-vision-run-2-train-rollout-no-diverge-iter_0000079-complete	0.5840	0.580	292/500	13 (2)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-9b-0918-i79/
medcli-v2-grpo-9b-step49	Qwen3.5-9B-Q38-MedCLI-V2-CMT-EHRDQ-GRPO-Step-49-Epoch-1	0.5560	0.548	278/500	32 (4)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-9b-step49/
medcli-v2-grpo-4b-scratch-i41	20261012-medcli-v2-grpo-qwen3.5-4b-q38-ctm-top50-ehr-prec10-scratch-iter_0000041-complete	withheld: 1/500 unscored	0.494	248/499	40 (1)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-4b-scratch-i41/
medcli-v2-grpo-4b-step79	Qwen3.5-4B-Q38-MedCLI-V2-CMT-EHRDQ-GRPO-Step-79-Epoch-1p5	withheld: 2/500 unscored	0.498	250/498	23 (1)	TP2 x DP4, 262144	qwen3_coder	True	medcli-v2-grpo-4b-step79/

Each model directory holds independent .tar.zst shards of <run-id>/<alias>/ (the smoke and full jobs: per-trial results, agent logs and trajectories; serve logs, resource metrics, the corruption audit). Restore with bash <alias>/restore.sh <empty-dir>. reports/ has each model's results/benchmark.json.