20261018-reasoning-qa-eval-open-weights: reasoning and medical QA evaluation

Six checkpoints (MedCLI-V2 SFT 27B, GRPO 9B and 4B, and their Qwen bases) on medical transfer and general-capability retention benchmarks, from examples/opencompass_eval in ProRL-Agent-Server-v2 (OpenCompass c2d92a9b and simple-evals 652c89d, pinned).

Benchmark	Questions x completions	Max output tokens	Scoring
MedQA-US, 4 options (GBaker/MedQA-USMLE-4-options)	1,273 x 5	32,768	ANSWER: X letter
MedXpertQA-Text test	2,450 x 3	32,768	ANSWER: X letter (A-J)
AIME 2025 / AIME 2026	30 x 32 each	81,920	last \boxed{} = gold integer
GPQA Diamond (OpenCompass's fixed option order)	198 x 8	32,768	ANSWER: X letter
LiveCodeBench code generation, release_v6, 2024-08-01..2025-05-01	454 x 8	81,920	pass@1, tests run only to grade
HealthBench (Hard = its 1,000-conversation subset)	5,000 x 1	32,768	simple-evals rubric grading, GPT-5.4 (low)
HealthBench Professional	525 x 8	32,768	same, length-adjusted (2,000-char center, 0.0147 / 500 chars)
All six above again, 27Bs only ("(2x cap)" rows)	as above	2x each cap above	as above

Every model: SGLang 0.5.15, eight single-GPU replicas, temperature 1.0, top-p 0.95, top-k 20, the chat template's default thinking (Qwen3.8 template: reasoning_effort xhigh), no system prompt, no tools. Only the final answer (after the thinking) is graded; a completion that never closed its thinking scores zero. Scores are means over all completions; ± is the question-clustered standard error. HealthBench uses GPT-5.4 at low reasoning as the grader (HealthBench Professional's protocol; the original HealthBench release used GPT-4.1). * marks a stage graded only partly. For this run, HealthBench and HealthBench Professional ran on another node; their results are in hab-swe/20260923-reasoning-qa-eval-healthbench.

model	MedQA-US (4-opt)	MedXpertQA-Text	AIME 2025	AIME 2026	GPQA Diamond	LiveCodeBench v6
Apertus-8B-MeditronFO	49.4 ± 0.9	12.0 ± 0.4	0.0 ± 0.0	0.1 ± 0.1	24.3 ± 1.4	3.8 ± 0.6
Baichuan-M2-32B	89.3 ± 0.7	28.0 ± 0.8	7.5 ± 2.1	7.7 ± 1.9	65.3 ± 2.6	60.9 ± 2.0
Baichuan-M3-235B	92.8 ± 0.6	42.1 ± 0.9	87.3 ± 5.2	90.6 ± 4.3	74.8 ± 2.5	72.4 ± 1.8
ClinSeek-35B-A3B	89.7 ± 0.7	29.5 ± 0.8	88.8 ± 4.6	88.5 ± 4.4	80.5 ± 2.3	65.0 ± 1.8
Gemma-3-27B-MeditronFO-sglang	76.7 ± 0.9	18.0 ± 0.6	23.0 ± 6.7	21.2 ± 5.9	46.8 ± 2.6	28.9 ± 2.0
HuatuoGPT-3-27B	94.7 ± 0.6	49.4 ± 0.9	95.3 ± 2.1	96.5 ± 2.0	88.6 ± 2.0	86.8 ± 1.3
HuatuoGPT-3-9B	93.1 ± 0.6	38.6 ± 0.9	47.5 ± 7.4	52.2 ± 7.1	80.4 ± 2.5	50.8 ± 2.0
KumaKuma-Qwen3.5-4B-OPD-Experimental	90.5 ± 0.7	28.4 ± 0.8	76.6 ± 6.1	81.4 ± 5.6	76.7 ± 2.5	58.3 ± 2.0
KumaKuma-Qwen3.5-9B-R09	92.8 ± 0.6	34.9 ± 0.9	86.9 ± 4.9	87.8 ± 4.8	80.8 ± 2.3	68.1 ± 1.9
Lingshu-32B	73.8 ± 0.9	20.3 ± 0.6	6.5 ± 2.4	6.1 ± 2.4	37.9 ± 2.1	17.8 ± 1.5
Lingshu-7B	55.7 ± 1.0	14.6 ± 0.5	0.0 ± 0.0	0.2 ± 0.1	27.5 ± 1.5	6.2 ± 0.9
medgemma-1.5-4b-it	70.8 ± 1.0	17.1 ± 0.6	0.5 ± 0.5	0.8 ± 0.6	21.5 ± 1.6	6.8 ± 0.9
medgemma-27b-it	82.1 ± 0.9	21.9 ± 0.7	23.5 ± 6.5	20.8 ± 5.5	45.3 ± 2.5	26.2 ± 1.9
MediPhi-Instruct	42.7 ± 1.0	8.6 ± 0.4	0.1 ± 0.1	0.1 ± 0.1	25.8 ± 1.6	6.9 ± 0.9
Qwen3.6-35B-A3B	93.7 ± 0.6	41.9 ± 0.9	92.6 ± 4.5	92.4 ± 3.7	81.8 ± 2.2	78.6 ± 1.6
tmax-27b-complete	94.4 ± 0.6	44.4 ± 0.9	93.2 ± 4.6	93.6 ± 3.6	84.1 ± 2.3	83.4 ± 1.5
tmax-4b-complete	91.0 ± 0.7	28.6 ± 0.8	84.1 ± 5.3	86.8 ± 5.3	78.2 ± 2.5	66.7 ± 1.9
tmax-9b-complete	93.3 ± 0.6	35.8 ± 0.9	89.4 ± 4.7	89.0 ± 4.5	82.3 ± 2.3	72.3 ± 1.8

Layout: <checkpoint>/<stage>/ holds independent .tar.zst shards of that model's stage directory (opencompass/main/: predictions, per-sample results, usage logs, serve log; healthbench*/: responses, per-rubric grades, result). Restore with bash <checkpoint>/<stage>/restore.sh <empty-dir>. reports/<checkpoint>/<stage>.json holds each stage's summary.