LMSYS 发布开源聊天模型 Vicuna-13B,用 ShareGPT 对话微调 LLaMA,训练成本约 $300
TL;DR · AI 摘要
LMSYS åå¸å¼æºè天模å Vicuna-13Bï¼ç¨ ShareGPT 对è¯å¾®è° LLaMAï¼è®ç»ææ¬çº¦ $300 · AIHOT LMSYSï¼Blog...
核心要点
- 主题聚焦:LMSYS 发布开源聊天模型 Vicuna-13B,用 ShareGPT 对话微调 LLaMA,
- 来源:AI HOT 精选,建议结合原文判断细节。
- AI 分析暂不可用,本条为保底评分与摘要。
LMSYS åå¸å¼æºè天模å Vicuna-13Bï¼ç¨ ShareGPT 对è¯å¾®è° LLaMAï¼è®ç»ææ¬çº¦ $300 · AIHOT
LMSYSï¼Blog
·
2026-10-03 23:50
1 å°æ¶å
ç²¾é
AI è¯å
79
LMSYS åå¸å¼æºè天模å Vicuna-13Bï¼ç¨ ShareGPT 对è¯å¾®è° LLaMAï¼è®ç»ææ¬çº¦ $300
Blog Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90% ChatGPT Quality We introduce Vicuna-13B, an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. Preliminary evaluation using GPT-4 as a judge shows Vicuna-13B achiev... The Vicuna Team March 30, 2023
AI 导读
LMSYS å¢éåå¸ Vicuna-13Bï¼éè¿çº¦ 70K ShareGPT ç¨æ·å ±äº«å¯¹è¯å¾®è° LLaMAï¼è®ç»ææ¬çº¦ $300ï¼ä»£ç ãæéåå¨çº¿ demo 以éåä¸è®¸å¯å ¬å¼ãGPT-4 ä½ä¸ºè¯å®¡ç忥è¯ä¼°æ¾ç¤ºå ¶è¾¾å° ChatGPT/Bard 90% 以ä¸è´¨éï¼å¨ 45% é®é¢ä¸ä¸éäº ChatGPTï¼å¢éåæ¶æåºåºäº GPT-4 çèªå¨è¯æµæ¡æ¶ï¼å¹¶è¯´æå ¶å°éä¸¥è°¨æ¹æ³ã
æ¨èçç±
åæç»åºè®ç»æ°æ®ãææ¬å GPT-4 è¯å®¡æ¹æ³çç»èï¼è¯»è 坿®æ¤äºè§£æ©æå¼æºå¯¹è¯æ¨¡åçå¤ç°è·¯å¾ä¸è¯æµæè·¯ã
æ£æ · AI ç¿»è¯
æä»¬æ¨åº Vicuna-13Bï¼ä¸ä¸ªå¼æºè天æºå¨äººï¼éè¿å¨ ShareGPT æ¶éçç¨æ·å ±äº«å¯¹è¯ä¸å¾®è° LLaMA è®ç»èæãä½¿ç¨ GPT-4 ä½ä¸ºè¯å¤ç忥è¯ä¼°æ¾ç¤ºï¼Vicuna-13B è¾¾å°äº OpenAI ChatGPT å Google Bard è¶ è¿ 90%* çè´¨éï¼åæ¶å¨è¶ è¿ 90% * çæ åµä¸ä¼äº LLaMA å Stanford Alpaca çå ¶ä»æ¨¡åãè®ç» Vicuna-13B çææ¬çº¦ä¸º 300 ç¾å ã 代ç å æé ï¼ä»¥åå¨çº¿ æ¼ç¤º ï¼åå·²å ¬å¼ä¾éåä¸ä½¿ç¨ã
Vicunaï¼ç± stable diffusion 2.1 çæï¼
*æ ¹æ®ä¸é¡¹æè¶£ä¸éç§å¦ç GPT-4 è¯ä¼°ãéè¦è¿ä¸æ¥ä¸¥æ ¼è¯ä¼°ã
Vicuna æå¤å¥½ï¼
å¨ç¨ 70K ç¨æ·å ±äº«ç ChatGPT 对è¯å¾®è° Vicuna åï¼æä»¬åç°ä¸ Alpaca ç¸æ¯ï¼Vicuna è½å¤çææ´è¯¦ç»ãç»ææ´æ¸ æ°ççæ¡ï¼è§ä¸æ¹ç¤ºä¾ï¼ï¼è´¨éä¸ ChatGPT ç¸å½ã
ç¶èï¼è¯ä¼°è天æºå¨äººä»æ¥é½ä¸æ¯ä¸é¡¹ç®åçä»»å¡ã éç GPT-4 çææ°è¿å±ï¼æä»¬å¥½å¥å ¶è½åæ¯å¦å·²è¾¾å°ç±»äººæ°´å¹³ï¼ä»èè½å¤å®ç°ç¨äºåºåçæåæ§è½è¯ä¼°çèªå¨åè¯ä¼°æ¡æ¶ã æä»¬ç忥åç°è¡¨æï¼å¨æ¯è¾è天æºå¨äººççæ¡æ¶ï¼GPT-4 å¯ä»¥äº§çé«åº¦ä¸è´çæåå详ç»è¯ä¼°ï¼è§ä¸æ¹ GPT-4 å¤æç¤ºä¾ï¼ã åºäº GPT-4 ç忥è¯ä¼°æ»ç»å¨å¾ 1 ä¸ï¼æ¾ç¤º Vicuna è¾¾å°äº Bard/ChatGPT 90% * çè½åã è½ç¶è¿ä¸ªæåºçæ¡æ¶æ¾ç¤ºåºèªå¨åè天æºå¨äººè¯ä¼°çæ½åï¼ ä½å®è¿ä¸æ¯ä¸ç§ä¸¥æ ¼çæ¹æ³ ã æå»ºè天æºå¨äººè¯ä¼°ç³»ç»ä»ç¶æ¯ä¸ä¸ªéè¦è¿ä¸æ¥ç ç©¶ç弿¾é®é¢ãæ´å¤ç»èå¨è¯ä¼°é¨åæä¾ã
å¾ 1. ç± GPT-4* è¯ä¼°çç¸å¯¹ååºè´¨é
å¨çº¿æ¼ç¤º
å¨ è¿é è¯ç¨ Vicuna-13B æ¼ç¤ºï¼
æ¦è¿°
大åè¯è¨æ¨¡åï¼LLMï¼çå¿«éåå±å·²ç»å½»åºæ¹åäºè天æºå¨äººç³»ç»ï¼å¸¦æ¥äºåææªæçæºè½æ°´å¹³ï¼æ£å¦ OpenAI ç ChatGPT æç¤ºãç¶èï¼å°½ç®¡å ¶æ§è½ä»¤äººå°è±¡æ·±å»ï¼ChatGPT çè®ç»åæ¶æç»èä»ä¸æ¸ æ¥ï¼é»ç¢äºè¯¥é¢åçç ç©¶å弿ºåæ°ãå Meta LLaMA å Stanford Alpaca 项ç®çå¯åï¼æä»¬æ¨åº Vicuna-13Bï¼ä¸ä¸ªç±å¢å¼ºæ°æ®éåæäºä½¿ç¨ã坿©å±çåºç¡è®¾æ½æ¯æç弿ºè天æºå¨äººãéè¿å¨ ShareGPT.com æ¶éçç¨æ·å ±äº«å¯¹è¯ä¸å¾®è° LLaMA åºç¡æ¨¡åï¼Vicuna-13B å·²å±ç°åºä¸ Stanford Alpaca çå ¶ä»å¼æºæ¨¡åç¸æ¯å ·æç«äºåçæ§è½ãè¿ç¯å客æç« æä¾äº Vicuna-13B æ§è½ç忥è¯ä¼°ï¼å¹¶æè¿°äºå ¶è®ç»åæå¡åºç¡è®¾æ½ãæä»¬è¿é请社åºä¸æä»¬çå¨çº¿æ¼ç¤ºäºå¨ï¼ä»¥æµè¯è¿ä¸ªè天æºå¨äººçè½åã
å¾ 2. 工使µç¨æ¦è¿°
å¾ 2 æ¦è¿°äºæä»¬çå·¥ä½ãé¦å ï¼æä»¬ä» ShareGPT.com æ¶éäºçº¦ 70K 对è¯ï¼è¿æ¯ä¸ä¸ªç¨æ·å¯ä»¥åäº«å ¶ ChatGPT 对è¯çç½ç«ãæ¥ä¸æ¥ï¼æä»¬å¢å¼ºäº Alpaca æä¾çè®ç»èæ¬ï¼ä»¥æ´å¥½å°å¤çå¤è½®å¯¹è¯åé¿åºåãè®ç»å¨ 8 个 A100 GPU ä¸ä½¿ç¨ PyTorch FSDP å¨ä¸å¤©å 宿ãä¸ºäºæä¾æ¼ç¤ºæå¡ï¼æä»¬å®ç°äºä¸ä¸ªè½»é级çåå¸å¼æå¡ç³»ç»ãæä»¬éè¿å建ä¸ç» 80 ä¸ªå¤æ ·åçé®é¢å¹¶å©ç¨ GPT-4 æ¥è¯å¤æ¨¡åè¾åºæ¥å¯¹æ¨¡åè´¨éè¿è¡äºåæ¥è¯ä¼°ãä¸ºäºæ¯è¾ä¸¤ä¸ªä¸åçæ¨¡åï¼æä»¬å°æ¯ä¸ªæ¨¡åçè¾åºåå¹¶å°æ¯ä¸ªé®é¢çå个æç¤ºä¸ãç¶åå°æç¤ºåéç» GPT-4ï¼ç±å ¶è¯ä¼°åªä¸ªæ¨¡åæä¾äºæ´å¥½çååºãLLaMAãAlpacaãChatGPT å Vicuna çè¯¦ç»æ¯è¾å¦ä¸è¡¨ 1 æç¤ºã
表 1. å 个è忍¡åä¹é´çæ¯è¾
模ååç§°
LLaMA
Alpaca
Vicuna
Bard/ChatGPT
æ°æ®é
å ¬å¼å¯ç¨çæ°æ®é
(1T token)
æ¥èª davinci-003 API çèªæå¯¼
(52K æ ·æ¬)
ç¨æ·å ±äº«ç对è¯
(70K æ ·æ¬)
ä¸éç¨
è®ç»ä»£ç
å¯ç¨
è¯ä¼°ææ
妿¯åºå
ä½è è¯ä¼°
GPT-4 è¯ä¼°
æ··å
è®ç»ææ¬
(7B)
82K GPU å°æ¶
$500 (æ°æ®) + $100 (è®ç»)
$140 (è®ç»)
(13B)
135K GPU å°æ¶
$300 (è®ç»)
è®ç»
Vicuna æ¯éè¿ä½¿ç¨ä» ShareGPT.com éè¿å ¬å ± API æ¶éç大约 70K ç¨æ·å ±äº«å¯¹è¯å¯¹ LLaMA åºç¡æ¨¡åè¿è¡å¾®è°èå建çã为äºç¡®ä¿æ°æ®è´¨éï¼æä»¬å° HTML 转æ¢å markdown å¹¶è¿æ»¤æä¸äºä¸é彿ä½è´¨éçæ ·æ¬ãæ¤å¤ï¼æä»¬å°åé¿ç对è¯åæé忍¡åæå¤§ä¸ä¸æé¿åº¦çè¾å°ç段ã
æä»¬çè®ç»é æ¹å»ºç«å¨ Stanford ç alpaca ä¹ä¸ï¼å¹¶è¿è¡äºä»¥ä¸æ¹è¿ã
- å¤è½®å¯¹è¯ï¼ æä»¬è°æ´è®ç»æå¤±ä»¥èèå¤è½®å¯¹è¯ï¼å¹¶ä» æ ¹æ®è天æºå¨äººçè¾åºè®¡ç®å¾®è°æå¤±ã
- å åä¼åï¼ ä¸ºäºä½¿ Vicuna è½å¤çè§£é¿ä¸ä¸æï¼æä»¬å°æå¤§ä¸ä¸æé¿åº¦ä» alpaca ä¸ç 512 æ©å±å° 2048ï¼è¿å¤§å¤§å¢å äº GPU å åéæ±ãæä»¬éè¿å©ç¨ æ¢¯åº¦æ£æ¥ç¹ å éªå注æå æ¥åºå¯¹å åååã
- éè¿ç«ä»·å®ä¾é使æ¬ï¼ è®ç»æ°æ®é大 40 åä¸åºåé¿åº¦é¿ 4 åï¼è¿å¯¹è®ç»è´¹ç¨ææäºç¸å½å¤§çææãæä»¬éç¨ SkyPilot æç®¡ç«ä»·å®ä¾ ï¼éè¿å©ç¨æ´ä¾¿å®çç«ä»·å®ä¾ä»¥åé对æ¢å çèªå¨æ¢å¤åèªå¨åºå忢æ¥é使æ¬ãè¯¥è§£å³æ¹æ¡å°è®ç» 7B 模åçææ¬ä» $500 éä½å°çº¦ $140 ï¼å° 13B 模åçææ¬ä»çº¦ $1K éä½å° $300 ã
æå¡
æä»¬æå»ºäºä¸ä¸ªæå¡ç³»ç»ï¼è½å¤éè¿åå¸å¼å·¥ä½å¨ä¸ºå¤ä¸ªæ¨¡åæä¾æå¡ã宿¯æçµæ´»å°æå ¥æ¥èªæ¬å°é群åäºç GPU å·¥ä½å¨ãéè¿å©ç¨ SkyPilot ä¸ç容鿧å¶å¨åæç®¡ç«ä»·å®ä¾åè½ï¼è¯¥æå¡ç³»ç»å¯ä»¥å¾å¥½å°ä¸æ¥èªå¤ä¸ªäºçæ´ä¾¿å®çç«ä»·å®ä¾é å使ç¨ï¼ä»¥é使塿æ¬ãå®ç®åæ¯ä¸ä¸ªè½»é级å®ç°ï¼æä»¬æ£å¨åªåå°æ´å¤ææ°ç ç ç©¶ éæå°å ¶ä¸ã
å¦ä½è¯ä¼°è天æºå¨äººï¼
è¯ä¼° AI è天æºå¨äººæ¯ä¸é¡¹å ·ææææ§çä»»å¡ï¼å 为å®éè¦èå¯è¯è¨çè§£ãæ¨çåä¸ä¸ææç¥è½åãéç AI è天æºå¨äººåå¾è¶æ¥è¶å è¿ï¼å½åç弿ºåºåå¯è½å·²ä¸åè¶³å¤ãä¾å¦ï¼æ¯å¦ç¦ Alpaca ä¸ä½¿ç¨çè¯ä¼°æ°æ®é self-instruct ï¼SOTA è天æºå¨äººå¯ä»¥ææå°åçï¼ä½¿å¾äººç±»é¾ä»¥è¾¨å«æ§è½å·®å¼ãæ´å¤å±éå æ¬è®ç»/æµè¯æ°æ®æ±¡æä»¥åå建æ°åºåçæ½å¨é«ææ¬ã为äºè§£å³è¿äºé®é¢ï¼æä»¬æåºäºä¸ä¸ªåºäº GPT-4 çè¯ä¼°æ¡æ¶ï¼ä»¥èªå¨åè天æºå¨äººæ§è½è¯ä¼°ã
é¦å ï¼æä»¬è®¾è®¡äºå «ä¸ªé®é¢ç±»å«ï¼ä¾å¦è´¹ç±³é®é¢ãè§è²æ®æ¼åºæ¯åç¼ç /æ°å¦ä»»å¡ï¼ä»¥æµè¯è天æºå¨äººæ§è½çå个æ¹é¢ãéè¿ç²¾å¿çæç¤ºå·¥ç¨ï¼GPT-4 è½å¤çæå¤æ ·ä¸å ·ææææ§çé®é¢ï¼åºçº¿æ¨¡åé¾ä»¥åºå¯¹ãæä»¬æ¯ä¸ªç±»å«éæ©å个é®é¢ï¼å¹¶ä»äºä¸ªè天æºå¨äººæ¶éçæ¡ï¼LLaMAãAlpacaãChatGPTãBard å Vicunaãç¶åæä»¬è¦æ± GPT-4 æ ¹æ®æç¨æ§ãç¸å ³æ§ãåç¡®æ§åç»èå¯¹å ¶çæ¡è´¨éè¿è¡è¯åãæä»¬åç° GPT-4 ä¸ä» è½äº§çç¸å¯¹ä¸è´çåæ°ï¼è¿è½è¯¦ç»è§£é为ä»ä¹ç»åºè¿æ ·çåæ°ï¼è¯¦ç»ç¤ºä¾ 龿¥ ï¼ãç¶èï¼æä»¬ä¹æ³¨æå° GPT-4 ä¸å¤ªæ é¿è¯å¤ç¼ç /æ°å¦ä»»å¡ã
å¾ 3. GPT-4 è¯ä¼°çååºæ¯è¾
å¾ 3 æ¾ç¤ºäºææåºçº¿ä¸ Vicuna ä¹é´çæ¯è¾ç»æãGPT-4 å¨è¶ è¿ 90% çé®é¢ä¸æ´å好 Vicuna èéæå è¿ç弿ºæ¨¡åï¼LLaMAãAlpacaï¼ï¼å¹¶ä¸å®ç¸å¯¹äºä¸ææ¨¡åï¼ChatGPTãBardï¼åå¾äºæç«äºåçæ§è½ãå¨ 45% çé®é¢ä¸ï¼GPT-4 å° Vicuna çååºè¯ä¸ºä¼äºæçäº ChatGPT çã ç±äº GPT-4 对æ¯ä¸ªååºæ 10 åå¶åé å®éåæ°ï¼æä»¬éè¿å°æ¯ä¸ªæ¨¡åå¨ 80 个é®é¢ä¸è·å¾çåæ°ç¸å ï¼è®¡ç®æ¯ä¸ªï¼åºçº¿ï¼Vicunaï¼æ¯è¾å¯¹çæ»åãå¦è¡¨ 2 æç¤ºï¼Vicuna çæ»åæ¯ ChatGPT ç 92%ã尽管æè¿åå¾äºè¿å±ï¼è¿äºè天æºå¨äººä»ç¶é¢ä¸´å±éï¼ä¾å¦é¾ä»¥å¤çåºæ¬æ°å¦é®é¢æç¼ç è½åæéã
表 2. GPT-4 è¯ä¼°çæ»åã
åºçº¿
åºçº¿åæ°
Vicuna åæ°
LLaMA-13B
513.0
694.0
Alpaca-13B
583.0
704.0
Bard
664.0
655.5
ChatGPT
693.0
638.0
è½ç¶è¿ä¸ªæåºçè¯ä¼°æ¡æ¶å±ç¤ºäºè¯ä¼°è天æºå¨äººçæ½åï¼ä½å®è¿ä¸æ¯ä¸ç§ä¸¥è°¨ææççæ¹æ³ï¼å 为大åè¯è¨æ¨¡å容æäº§çå¹»è§ã为è天æºå¨äººå¼åä¸ä¸ªå ¨é¢ãæ ååçè¯ä¼°ç³»ç»ä»ç¶æ¯ä¸ä¸ªéè¦è¿ä¸æ¥ç ç©¶ç弿¾é®é¢ã
ç¼è¾ ï¼å¨è¿ç¯å客æç« ä¹åï¼æä»¬å¯¹è¿ç§åºäº GPT4 çè¯ä¼°æ¹æ³è¿è¡äºæ´æ·±å ¥çç ç©¶ãæ¬¢è¿é 读æä»¬çæ°è®ºæ Judging LLM-as-a-judge å¹¶å°è¯æ°çè¯ä¼° å·¥å · ã
å±éæ§
æä»¬æ³¨æå°ï¼ä¸å ¶ä»å¤§åè¯è¨æ¨¡å类似ï¼Vicuna ä¹å卿äºå±éæ§ãä¾å¦ï¼å®ä¸æ é¿æ¶åæ¨çææ°å¦çä»»å¡ï¼å¹¶ä¸å¨åç¡®è¯å«èªèº«æç¡®ä¿å ¶è¾åºçäºå®åç¡®æ§æ¹é¢å¯è½åå¨å±éãæ¤å¤ï¼å®å°æªç»è¿å åä¼å以ä¿è¯å®å ¨æ§æåè½»æ½å¨çæ¯æ§æåè§ã为äºè§£å³å®å ¨é®é¢ï¼æä»¬å¨å¨çº¿æ¼ç¤ºä¸ä½¿ç¨ OpenAI moderation API æ¥è¿æ»¤ä¸éå½çç¨æ·è¾å ¥ãå°½ç®¡å¦æ¤ï¼æä»¬é¢è®¡ Vicuna å¯ä»¥ä½ä¸ºæªæ¥ç ç©¶è§£å³è¿äºå±éæ§ç弿¾èµ·ç¹ã
åå¸
卿们ç馿¬¡åå¸ä¸ï¼æä»¬å°å¨ GitHub ä»åºå享è®ç»ãæå¡åè¯ä¼°ä»£ç ï¼ https://github.com/lm-sys/FastChat ã æä»¬è¿åå¸äº Vicuna-13B æ¨¡å æé ã æ²¡æè®¡åå叿°æ®éãå å ¥æä»¬ç Discord æå¡å¨å¹¶å ³æ³¨æä»¬ç Twitter 以è·åææ°æ´æ°ã
许å¯
å¨çº¿æ¼ç¤ºæ¯ä» ééåä¸ç¨éçç ç©¶é¢è§ï¼å LLaMA çæ¨¡å è®¸å¯ ãOpenAI çææ°æ®ç ä½¿ç¨æ¡æ¬¾ 以å ShareGPT ç éç§å®è·µ 约æã妿æ¨åç°ä»»ä½æ½å¨è¿è§ï¼è¯·èç³»æä»¬ã 代ç å¨ Apache License 2.0 ä¸åå¸ã
è´è°¢
æä»¬è¦æè°¢ BAIR ç Xinyang GengãHao Liu å Eric Wallaceï¼æ¯å¦ç¦ Alpaca å¢éç Xuecheng Li å Tianyi Zhang çæ·±å ¥è®¨è®ºååé¦ï¼MBZUAI ç Qirong Ho 卿å¡éç¾¤ä¸æä¾çæ¯æã请æ¥ç BAIR å ³äºä»ä»¬è天æºå¨äºº Koala å¹¶è¡å·¥ä½çå客æç« ã
å¢é
è¿æ¯ä¸æ¥èªå¤ä¸ªæºæçåä½è çèååªåï¼å æ¬ UC BerkeleyãCMUãæ¯å¦ç¦ãUC San Diego å MBZUAIã
- å¦çï¼æåæ¯é¡ºåºï¼ï¼ Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang (â), Lianmin Zheng (â), Siyuan Zhuang, Yonghao Zhuang
- 顾é®ï¼æåæ¯é¡ºåºï¼ï¼ Joseph E. Gonzalez, Ion Stoica, Eric P. Xing
â é讯ä½è ï¼ Lianmin Zheng ( [email protected] ), Hao Zhang ( [email protected] ), æ LMSYS ( [email protected] )ã
å¼ç¨
@misc{vicuna2023,
title = {Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90\%* ChatGPT Quality},
url = {https://lmsys.org/blog/2023-03-30-vicuna/},
author = {Chiang, Wei-Lin and Li, Zhuohan and Lin, Zi and Sheng, Ying and Wu, Zhanghao and Zhang, Hao and Zheng, Lianmin and Zhuang, Siyuan and Zhuang, Yonghao and Gonzalez, Joseph E. and Stoica, Ion and Xing, Eric P.},
month = {March},
year = {2023}
}å¨è¿ç¯å客æç« ä¹åï¼æä»¬æ©å±äºåºäº GPT-4 çè¯ä¼°æ³æ³ï¼å¹¶æ°åäºä¸ç¯æ´æ£å¼ç论æï¼ç³»ç»å°ç ç©¶äºè¿ç§âLLM-as-a-judgeâæ¹æ³ã 欢è¿é 读并å¼ç¨è¿ç¯è®ºæï¼ Judging LLM-as-a-judge with MT-Bench and Chatbot Arena ã
æ¥æºï¼ LMSYSï¼Blog · lmsys.org
弿ºçæ
模ååå¸
è¯æµåºå
#
è¯æµ/åºå