å»çãã¡ã€ã³ç¹åLLMã®æ§èœã¯ã©ããã£ãŠè©äŸ¡ããïŒ
BioMistral ã®è«æããšãŠãè¯ããµãŒãã€ãå ŒããŠããã®ã§ããã£ãããªã®ã§ç°¡åã«ãŸãšããŠã¿ãããšã«ããŸãããéãªã®ã§ãè²ã ééãã¯ãããããããŸããã
ãªãŒãã³ãªã¢ãã«
BioGPT (2022)
BioGPT 㯠GPT ç³»åã®ã¢ãã«ã§ããèŠæš¡æã¯ GPT-2 ãšåãã§ãGPT-2 㯠LLM ãšåŒã¹ããã¯çã æªããã§ãããdecoder ç³»ã¢ãã«ã®å ç¥ãšããŠåãäžããŸãã
PubMed ã䜿ã£ãŠãã«ã¹ã¯ã©ããåŠç¿ããŠããŸãã
è©äŸ¡ã¯ BC5CDR ã DDI ã®ãã㪠Relation Extraction ç³»ã®ããŒã¿ã»ãããHoC ã®ãããªææžåé¡ç³»ã®ããŒã¿ã»ãããããã« PubMedQA ã䜿ã£ãŠããŸãã
ããšãé¢çœãã®ããäœãå°éçšèªãå ¥ããŠãæ®éã®ãã¡ã€ã³ã® GPT-2 ãšããã¹ãçæãããæã®éããèŠæ¯ã¹ãŠããŸãããã®æä»£ã®èšèªã¢ãã«ã¯ç¶ããäºæž¬ãããããšããã§ããªãã£ãã®ã§ãããããããšãè¡ãããŠããã®ã§ãã
ClinicalGPT (2023)
ClinicalGPT 㯠BLOOM-7B ãããŒã¹ã«ãªã£ãŠããŸããéåžžãã¡ã€ã³ã® LLM ã§ãã BLOOM-7B ã«å¯ŸããŠãå»çãã¡ã€ã³ã® QA ã蚺å¯ããã«ãã¿ãŒã³äŒè©±ã®ãããªããŒã¿ã»ããã§ Supervised Fine-tuning ããããšã人éã®ã¹ã³ã¢ä»ãã§ PPO ãã RLHF ããã£ãŠããããã§ãïŒãã®äººéãå»è ãã©ããã¯ããããããªãã£ãïŒã
ãªããäžåœèªã®ã¢ãã«ã§ãã
è©äŸ¡ã¯ãäžåœèªã®ã¢ãã«ãªã®ã§ãããããããŸããã§ããããäž»ã«ã
- äŒè©±ããŒã¿ã»ããã䜿ã£ãŠãåç §æãšã©ã®ããã䌌ãŠãããã BLEU, ROUGE ã§è©äŸ¡
- QAããŒã¿ã»ããã§æ£è§£ã§ããããè©äŸ¡
ã®2éãããã£ãŠããæããããŸãã
PMC-LLaMA (2023)
PMC-LLaMA 㯠Llama ããŒã¹ã®ã¢ãã«ã§ããClinicalGPT ãšã®éããšããŠãè«æãå»åŠç³»ææžã§è¿œå äºååŠç¿ãè¡ã£ãŠããŸãã
è©äŸ¡ããŒã¿ã»ãã㯠MedQA (USMLE)ãMedMCQAãPubMedQA ã® 3 ã€ã§ãã
MedAlpaca (2023)
ããã Llama ããŒã¹ã§ããèšç·ŽããŒã¿åéã®ããã«ãWeb äžã®ããŒã¿ãªã©ãããå»çé¢ä¿ã®ãã®ãããéããŠããã®ãç¹åŸŽã§ãã
è©äŸ¡ããŒã¿ã»ãã㯠MedQA (USMLE) ã ãã§ãã
ChatDoctor (2023)
ããã Llama ããŒã¹ã§ããhealthcaremagic ãšãããªã³ã©ã€ã³ã®èšºçãµã€ãããäŒè©±ãéã㊠instruction tuning ããŠããŸãã
è©äŸ¡ã¯ãªããšãChatGPT ãæ£è§£ããŒã¿ãšã¿ãªããäžã§ãChatGPT ã®åºåããã¹ããšã®é¡äŒŒåºŠã BERTScore ã§ïŒïŒïŒèšç®ããŠããŸããããã¯å€§èã
BioMedGPT-LM-7B (2023)
BioMedGPT-LM-7B 㯠Llama 2 ããŒã¹ã®ã¢ãã«ã§ãå»åŠè«æã§è¿œå äºååŠç¿ããŠããŸãã
ããã«ãå»åŠãšã¯ããå€ããŸãããBioMedGPT-LM-7B ãããŒã¹ã«ãååæ§é ãåŠã¶ Molecule Encoder ãšããã®ãšã¿ã³ãã¯è³ªæ§é ãåŠã¶ Protein Encoder ãšããã®ãä»ãå ãã BioMedGPT-10B ãšããã®ãŸã§äœã£ãŠããŸãã
è©äŸ¡ããŒã¿ã»ããã®ãã¡ãå»åŠç³»ã¿ã¹ã¯ã«ãããŠã¯ãMedQA (USMLE)ãMedMCQAãPubMedQA ã䜿ã£ãŠããŸãã
MediTron-7B, 70B (2023)
MediTron ãç»å Žããéã«ã¯çµæ§è©±é¡ã«ãªã£ãèšæ¶ããããŸããããã Llama 2 ããŒã¹ã®ã¢ãã«ã§ãã
ãšããããå³ããã£ã¡ãåãããããã
è©äŸ¡ããŒã¿ã»ãã㯠PubMedQAãMedMCQAãMedQA (USMLE) ã«å ããŠãMMLU-MedicalïŒMMLU ããŒã¿ã»ããã®ãã¡å»çã«ããŽãªã«å±ãããã®ã ãæœåºãããã®ïŒãåããŠç»å ŽããŸããã
BioMistral (2024)
BioMistral ã¯ååéã Mistral ããŒã¹ã§ãã
MMLU-MedicalãMedQA (USMLE)ãMedMCQAãPubMedQA ã® 4 ã€ã§è©äŸ¡ããŠããŸãã
ã¯ããŒãºãªã¢ãã«
Med-PALM (Google DeepMind) ã MedPrompt (Microsoft) ããããŸããã¯ããŒãºãªé å㯠Microsoft ãš Google ãããã³ããèžã§ããã¡ãã£ããããŠããã ããªã®ã§ã庶æ°ã«ã¯è¡ãæ«ãã¿ãŸãããããããŸãããã
éèŠãªã®ã¯ãè©äŸ¡ãã©ãããŠãããã ãšæããŸãã
Med-PALM 2 ã¯ãMedQA (USMLE)ãPubMedQAãMedMCQAãMMLU-Medical ã§è©äŸ¡ããŠããŸãïŒ+ å°éå®¶ã«ãã宿§è©äŸ¡ããããŸãïŒã
MedPrompt ããMedQA (USMLE)ãPubMedQAãMedMCQAãMMLU-Medical ã䜿ã£ãŠããŸããã
ãŸãšã
- 2024幎åé ã®çŸåšãå»çãã¡ã€ã³ç¹åLLMã®ãã³ãããŒã¯ãšããŠã¯ãMedQA (USMLE)ãPubMedQAãMedMCQAãMMLU-MedicalïŒMMLU ããŒã¿ã»ããã®ãã¡å»çã«ããŽãªã«å±ãããã®ã ãæœåºãããã®ïŒãäž»æµã§ããããšãåãããŸããã
- æ¥æ¬ã§ã¯ãMedQA (USMLE) ã®æ¥æ¬èªçã«çžåœãã IgakuQA ãšãMMLU ã®æ¥æ¬èªçã«çžåœãã JMMLU ãããããããããã䜿ã£ãŠ Japanese Language Model Financial Evaluation Harness çãªãã¡ã€ã³ç¹åãªãŒããŒããŒãããããšããããããªãšæããŸãããPFNãããã£ãŠãããŸãããïŒïŒïŒïŒ
Discussion