hab-swe/20261018-swebv-loopback-open-weights

SWE-bench Verified, gateway-only agent egress: open-weights medical and other 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/20261018-tb21-swebv-open-weights in ProRL-Agent-Server-v2, run id 20261018-swebv-loopback-open-weights). 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 (ignored by templates without them; Baichuan-M2 also thinking_mode on). SGLang 0.5.15.post1 on eight B200s, TP2 x DP4 at context 262144 where the checkpoint allows (columns below otherwise); 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
huatuogpt-3-27b	HuatuoGPT-3-27B	0.7520	0.748	376/500	5 (2)	TP2 x DP4, 262144	qwen3_coder	True	huatuogpt-3-27b/
huatuogpt-3-9b	HuatuoGPT-3-9B	withheld: 3/500 unscored	0.540	270/497	10 (0)	TP2 x DP4, 262144	qwen3_coder	True	huatuogpt-3-9b/
kumakuma-qwen3.5-9b-r09	KumaKuma-Qwen3.5-9B-R09	0.5880	0.584	294/500	10 (2)	TP2 x DP4, 262144	qwen3_coder	True	kumakuma-qwen3.5-9b-r09/
kumakuma-qwen3.5-4b-opd	KumaKuma-Qwen3.5-4B-OPD-Experimental	withheld: 1/500 unscored	0.496	248/499	27 (0)	TP2 x DP4, 262144	qwen3_coder	True	kumakuma-qwen3.5-4b-opd/
qwen3.6-35b-a3b	Qwen3.6-35B-A3B	0.7000	0.698	350/500	2 (1)	TP2 x DP4, 262144	qwen3_coder	True	qwen3.6-35b-a3b/
clinseek-35b-a3b	ClinSeek-35B-A3B	0.5640	0.562	282/500	7 (1)	TP2 x DP4, 262144	qwen3_coder	True	clinseek-35b-a3b/
tmax-27b	tmax-27b-complete	withheld: 1/500 unscored	0.738	369/499	0 (0)	TP2 x DP4, 262144	qwen3_coder	True	tmax-27b/
tmax-9b	tmax-9b-complete	withheld: 2/500 unscored	0.600	301/498	17 (1)	TP2 x DP4, 262144	qwen3_coder	True	tmax-9b/
tmax-4b	tmax-4b-complete	0.4000	0.400	200/500	38 (0)	TP2 x DP4, 262144	qwen3_coder	True	tmax-4b/
baichuan-m2-32b	Baichuan-M2-32B	withheld: 1/500 unscored	0.036	18/499	4 (0)	TP2 x DP4, 131072	hermes_lenient	False	baichuan-m2-32b/
medgemma-27b-it	medgemma-27b-it	0.0280	0.028	14/500	11 (0)	TP2 x DP4, 131072	hermes_lenient	True	medgemma-27b-it/
gemma-3-27b-meditronfo	Gemma-3-27B-MeditronFO-sglang	withheld: 1/500 unscored	0.032	16/499	3 (0)	TP2 x DP4, 131072	hermes_lenient	True	gemma-3-27b-meditronfo/
lingshu-32b	Lingshu-32B	withheld: 1/500 unscored	0.002	1/499	2 (0)	TP2 x DP4, 128000	hermes_lenient	True	lingshu-32b/
lingshu-7b	Lingshu-7B	0.0040	0.004	2/500	0 (0)	TP2 x DP4, 128000	hermes_lenient	True	lingshu-7b/
apertus-8b-meditronfo	Apertus-8B-MeditronFO	0.0040	0.004	2/500	0 (0)	TP2 x DP4, 65536	hermes_lenient	False	apertus-8b-meditronfo/
mediphi-instruct	MediPhi-Instruct	0.0020	0.002	1/500	0 (0)	TP2 x DP4, 131072	hermes_lenient	False	mediphi-instruct/

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.