hugo/healthbench-br
Viewer • Updated • 1.78k • 37
Protocolos Clínicos BR
Adapting Qwen2.5-14B to Brazilian SUS clinical guidelines: 2 benchmarks, the synthetic corpus, and 8 model checkpoints from the paper's ablations.
Note Benchmark — 1,780 paired true/false clinical assertions for precise factual recall.
hugo/pcdt-qa
Viewer • Updated • 178 • 29Note Benchmark — 890 open-ended clinical questions, judged by an LLM (GPT-4.1).
hugo/protocolos-clinicos-br
Viewer • Updated • 22.4k • 30Note Corpus — the 178 raw guidelines (default config) plus the 13 synthetic configs used for CPT.
hugo/protocolos-clinicos-br-rl-4gen-14b
Text Generation • Updated • 25Note Best model — CPT (4 generators) + GRPO, LoRA adapter. HB 83.9 / PCDT 85.4.
hugo/protocolos-clinicos-br-cpt-1gen-14b
Text Generation • 15B • Updated • 25Note CPT only, 1 generator (GPT-4.1-mini), full fine-tuning. HB 69.6 / PCDT 66.3.
hugo/protocolos-clinicos-br-cpt-4gen-14b
Text Generation • 15B • Updated • 26Note CPT only, 4 generators, full fine-tuning. HB 71.1 / PCDT 86.3 — beats frontier models on PCDT-QA.
hugo/protocolos-clinicos-br-rl-1gen-14b
Text Generation • Updated • 26Note CPT (1 generator) + GRPO, LoRA adapter. HB 75.6 / PCDT 65.4.
hugo/protocolos-clinicos-br-cpt_lora-1gen-14b
Text Generation • Updated • 26Note CPT, 1 generator, LoRA adapter (no full FT). HB 67.2 / PCDT 65.2.
hugo/protocolos-clinicos-br-cpt_lora-4gen-14b
Text Generation • Updated • 24Note CPT, 4 generators, LoRA adapter (no full FT). HB 70.3 / PCDT 77.1.
hugo/protocolos-clinicos-br-cpt_replay-1gen-14b
Text Generation • 15B • Updated • 24Note CPT, 1 generator, with FineWeb-EN replay (1:1 mix). HB 63.8 / PCDT 65.6.
hugo/protocolos-clinicos-br-base_rl-1gen-14b
Text Generation • Updated • 24Note RL only — GRPO directly on Qwen2.5-14B (no CPT), LoRA adapter. HB 64.2 / PCDT 29.4.