Available models:
\n\n","updatedAt":"2023-11-28T09:45:25.224Z","author":{"_id":"5e6a3d4ea9afd5125d9ec064","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1584020801691-noauth.jpeg","fullname":"Stefan Schweter","name":"stefan-it","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":3685,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.6906628608703613},"editors":["stefan-it"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1584020801691-noauth.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2311.16079","authors":[{"_id":"6565b230ba64f727ddbee00e","user":{"_id":"654b92086a49f6f6e0edbcd9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/654b92086a49f6f6e0edbcd9/1bWswk7EcL04mWwgx1TpA.jpeg","isPro":true,"fullname":"Zeming Chen - Eric","user":"zechen-nlp","type":"user"},"name":"Zeming Chen","status":"claimed_verified","statusLastChangedAt":"2023-11-29T14:54:56.373Z","hidden":false},{"_id":"6565b230ba64f727ddbee00f","user":{"_id":"648211d39c44e3422ca7e4f3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/648211d39c44e3422ca7e4f3/GtxJQRx0kjSiNnZ-wXrJ7.png","isPro":false,"fullname":"Alejandro Hernández Cano","user":"alehc","type":"user"},"name":"Alejandro Hernández Cano","status":"claimed_verified","statusLastChangedAt":"2023-12-08T10:20:02.901Z","hidden":false},{"_id":"6565b230ba64f727ddbee010","user":{"_id":"62a0803e3e7d1dda046d475a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/62a0803e3e7d1dda046d475a/Oi1qHSaX7_PU1pS8qaXKv.jpeg","isPro":false,"fullname":"Angelika Romanou","user":"angelika","type":"user"},"name":"Angelika Romanou","status":"claimed_verified","statusLastChangedAt":"2024-12-02T09:05:56.077Z","hidden":false},{"_id":"6565b230ba64f727ddbee011","user":{"_id":"64bc12cf140491ca9f341c8f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64bc12cf140491ca9f341c8f/MZ3vilsnnO5jSeao4QNd-.jpeg","isPro":false,"fullname":"Antoine Bonnet","user":"AGBonnet","type":"user"},"name":"Antoine Bonnet","status":"claimed_verified","statusLastChangedAt":"2023-12-04T08:03:07.555Z","hidden":false},{"_id":"6565b230ba64f727ddbee012","name":"Kyle Matoba","hidden":false},{"_id":"6565b230ba64f727ddbee013","user":{"_id":"652c25ac2ecb5062d6c993bc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/652c25ac2ecb5062d6c993bc/DIZCtdyPBHVR9fwCNFSmC.jpeg","isPro":false,"fullname":"Francesco Salvi","user":"frasalvi","type":"user"},"name":"Francesco Salvi","status":"claimed_verified","statusLastChangedAt":"2023-11-29T14:54:59.987Z","hidden":false},{"_id":"6565b230ba64f727ddbee014","name":"Matteo Pagliardini","hidden":false},{"_id":"6565b230ba64f727ddbee015","user":{"_id":"60d3d1f16a42871838fdcfb3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/60d3d1f16a42871838fdcfb3/B6CqCaGT6Oq7s5txVOomg.jpeg","isPro":false,"fullname":"Simin Fan","user":"Olivia-umich","type":"user"},"name":"Simin Fan","status":"claimed_verified","statusLastChangedAt":"2024-03-14T10:07:00.343Z","hidden":false},{"_id":"6565b230ba64f727ddbee016","user":{"_id":"63b888106cf60e8c3d00dc86","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63b888106cf60e8c3d00dc86/lNCiIc424q60PC2D33vxN.jpeg","isPro":true,"fullname":"Andreas Köpf","user":"andreaskoepf","type":"user"},"name":"Andreas Köpf","status":"claimed_verified","statusLastChangedAt":"2026-01-15T15:10:03.719Z","hidden":false},{"_id":"6565b230ba64f727ddbee017","name":"Amirkeivan Mohtashami","hidden":false},{"_id":"6565b230ba64f727ddbee018","name":"Alexandre Sallinen","hidden":false},{"_id":"6565b230ba64f727ddbee019","name":"Alireza Sakhaeirad","hidden":false},{"_id":"6565b230ba64f727ddbee01a","name":"Vinitra Swamy","hidden":false},{"_id":"6565b230ba64f727ddbee01b","name":"Igor Krawczuk","hidden":false},{"_id":"6565b230ba64f727ddbee01c","name":"Deniz Bayazit","hidden":false},{"_id":"6565b230ba64f727ddbee01d","name":"Axel Marmet","hidden":false},{"_id":"6565b230ba64f727ddbee01e","name":"Syrielle Montariol","hidden":false},{"_id":"6565b230ba64f727ddbee01f","name":"Mary-Anne Hartley","hidden":false},{"_id":"6565b230ba64f727ddbee020","name":"Martin Jaggi","hidden":false},{"_id":"6565b230ba64f727ddbee021","user":{"_id":"654fb1c67490049d621516fb","avatarUrl":"/avatars/0f88177da236b1bce4a4a58fd0246884.svg","isPro":false,"fullname":"Antoine Bosselut","user":"atcbosselut","type":"user"},"name":"Antoine Bosselut","status":"claimed_verified","statusLastChangedAt":"2024-12-02T08:58:36.288Z","hidden":false}],"publishedAt":"2023-11-27T18:49:43.000Z","title":"MEDITRON-70B: Scaling Medical Pretraining for Large Language Models","summary":"Large language models (LLMs) can potentially democratize access to medical\nknowledge. While many efforts have been made to harness and improve LLMs'\nmedical knowledge and reasoning capacities, the resulting models are either\nclosed-source (e.g., PaLM, GPT-4) or limited in scale (<= 13B parameters),\nwhich restricts their abilities. In this work, we improve access to large-scale\nmedical LLMs by releasing MEDITRON: a suite of open-source LLMs with 7B and 70B\nparameters adapted to the medical domain. MEDITRON builds on Llama-2 (through\nour adaptation of Nvidia's Megatron-LM distributed trainer), and extends\npretraining on a comprehensively curated medical corpus, including selected\nPubMed articles, abstracts, and internationally-recognized medical guidelines.\nEvaluations using four major medical benchmarks show significant performance\ngains over several state-of-the-art baselines before and after task-specific\nfinetuning. Overall, MEDITRON achieves a 6% absolute performance gain over the\nbest public baseline in its parameter class and 3% over the strongest baseline\nwe finetuned from Llama-2. Compared to closed-source LLMs, MEDITRON-70B\noutperforms GPT-3.5 and Med-PaLM and is within 5% of GPT-4 and 10% of\nMed-PaLM-2. We release our code for curating the medical pretraining corpus and\nthe MEDITRON model weights to drive open-source development of more capable\nmedical LLMs.","upvotes":19,"discussionId":"6565b230ba64f727ddbee03a","githubRepo":"https://github.com/epfllm/meditron","githubRepoAddedBy":"auto","ai_summary":"MEDITRON, a suite of open-source LLMs for medical applications, achieves significant performance gains across benchmarks compared to both closed-source and smaller-scale models.","ai_keywords":["large language models","LLMs","medical knowledge","reasoning capacities","open-source","Llama-2","Nvidia's Megatron-LM","distributed trainer","medical corpus","PubMed articles","medical guidelines","medical benchmarks","task-specific finetuning","MEDITRON-70B","GPT-3.5","Med-PaLM","GPT-4","Med-PaLM-2","code curating","model weights"],"githubStars":2145},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"5e6a3d4ea9afd5125d9ec064","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1584020801691-noauth.jpeg","isPro":true,"fullname":"Stefan Schweter","user":"stefan-it","type":"user"},{"_id":"6032802e1f993496bc14d9e3","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6032802e1f993496bc14d9e3/w6hr-DEQot4VVkoyRIBiy.png","isPro":false,"fullname":"Omar Sanseviero","user":"osanseviero","type":"user"},{"_id":"5fcc1929563427b03e9af259","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1620642852901-5fcc1929563427b03e9af259.jpeg","isPro":false,"fullname":"David Adelani","user":"Davlan","type":"user"},{"_id":"5f0c746619cb630495b814fd","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1594651707950-noauth.jpeg","isPro":true,"fullname":"Lewis Tunstall","user":"lewtun","type":"user"},{"_id":"6201c0129dab2e6e083d023c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1656519120848-6201c0129dab2e6e083d023c.jpeg","isPro":false,"fullname":"Katie Link","user":"katielink","type":"user"},{"_id":"652c25ac2ecb5062d6c993bc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/652c25ac2ecb5062d6c993bc/DIZCtdyPBHVR9fwCNFSmC.jpeg","isPro":false,"fullname":"Francesco Salvi","user":"frasalvi","type":"user"},{"_id":"654a3b0be4ce07c7d77ca12f","avatarUrl":"/avatars/1cde762b02e81910899fe14471696c0c.svg","isPro":false,"fullname":"ANDRE COUTINHO CASTILLA","user":"castillamed","type":"user"},{"_id":"5e845f5d92fa4e3c26ddb32c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/_1LS-K5wSSDGwslAVQ3uF.png","isPro":false,"fullname":"Praise","user":"Praise2112","type":"user"},{"_id":"6281d941eeb15579946ca3ce","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6281d941eeb15579946ca3ce/0CdrBop_kjRkOqxUTYFbf.jpeg","isPro":false,"fullname":"Hui Sun","user":"CocoSun","type":"user"},{"_id":"64747f7e33192631bacd8831","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64747f7e33192631bacd8831/dstkZJ4sHJSeqLesV5cOC.jpeg","isPro":false,"fullname":"Taufiq Dwi Purnomo","user":"taufiqdp","type":"user"},{"_id":"64778e4b8ab7e732b6d7bb41","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/wprgE8gGM-fWJGXWOI92D.jpeg","isPro":false,"fullname":"siwei liu","user":"TedSiwei","type":"user"},{"_id":"654b92086a49f6f6e0edbcd9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/654b92086a49f6f6e0edbcd9/1bWswk7EcL04mWwgx1TpA.jpeg","isPro":true,"fullname":"Zeming Chen - Eric","user":"zechen-nlp","type":"user"}],"acceptLanguages":["*"]}">MEDITRON-70B: Scaling Medical Pretraining for Large Language Models
Abstract
MEDITRON, a suite of open-source LLMs for medical applications, achieves significant performance gains across benchmarks compared to both closed-source and smaller-scale models.
Large language models (LLMs) can potentially democratize access to medical knowledge. While many efforts have been made to harness and improve LLMs' medical knowledge and reasoning capacities, the resulting models are either closed-source (e.g., PaLM, GPT-4) or limited in scale (<= 13B parameters), which restricts their abilities. In this work, we improve access to large-scale medical LLMs by releasing MEDITRON: a suite of open-source LLMs with 7B and 70B parameters adapted to the medical domain. MEDITRON builds on Llama-2 (through our adaptation of Nvidia's Megatron-LM distributed trainer), and extends pretraining on a comprehensively curated medical corpus, including selected PubMed articles, abstracts, and internationally-recognized medical guidelines. Evaluations using four major medical benchmarks show significant performance gains over several state-of-the-art baselines before and after task-specific finetuning. Overall, MEDITRON achieves a 6% absolute performance gain over the best public baseline in its parameter class and 3% over the strongest baseline we finetuned from Llama-2. Compared to closed-source LLMs, MEDITRON-70B outperforms GPT-3.5 and Med-PaLM and is within 5% of GPT-4 and 10% of Med-PaLM-2. We release our code for curating the medical pretraining corpus and the MEDITRON model weights to drive open-source development of more capable medical LLMs.