Chat Vector ã®å¹æ vs. MoEãã¼ã¸ ã®å¹æ
- ååã®ç¶ãã§ãã次ã¯Chat Vectorå¦çã«ãã£ã¦ã©ããããæ¥æ¬èªãã£ããæ©è½ãæ¹åãã¦ããã®ããå ·ä½çã«ã¿ã¦ã¿ã¾ãã
- ä¸è¨ã®è¨äºã§ãChat Vectorã使ããªãã¦ãåç´ã«2ã¢ãã«ãMoEãã¼ã¸ããã ãã§ä¸å®ã®æ§è½åä¸ã観å¯ã§ãããã¨ã¯ç¢ºèªãã¦ãã¾ãã
- ã§ã¯ãChat Vectorãå ãããã¨ã§ã¢ãã«ã®ãã£ããæ§è½ã¯ã©ã®ããã«å½±é¿ãåããã®ã®ã§ããããã
åä½ã®7Bã¢ãã«ã«ãããChat Vectorã®å¹æ
- ã¾ããLightChatAssistant 2x7Bãã®åç¾éç¨ã§å¾ããããChatNTQ 7Bã«Mistral-7B-v0.2-Instructã®Chat Vectorãé©ç¨ããã¢ãã«ãã®æ§è½ã測å®ãã¦ã¿ã¾ãã
å称 | LightChat Assistant 2x7B |
Chat NTQ 7B random MoE |
ChatNTQ 7B +ChatVector |
ChatNTQ 7B |
ãã©ã¡ã¼ã¿ | 2x7B | 2x7B | 7B | 7B |
éååã¿ã¤ã | Q8_0 | Q8_0 | Q8_0 | Q8_0 |
ElyzaTasks100 ã¹ã³ã¢å¹³å |
3.31 | 3.25 | 3.23 | 3.06 |
æ¨æºèª¤å·® | 0.14 | 0.14 | 0.15 | 0.15 |
å¹³ååæ° | 273.61 | 261.27 | 280.04 | 240.46 |
- ChatVectorå¦çã ãã§å¹³åã¹ã³ã¢ã¯3.23ã«ä¸æãã¾ããã
- 以åä½ã£ããChatNTQ 7Bã®x2ã©ã³ãã ãã¼ã¸ãã®ã¹ã³ã¢ã3.25ã ã£ãã®ã§ããã®Chat Vectorã¯ã©ã³ãã ãã¼ã¸ã«è¿ãã¬ãã«ã®æ§è½ä¸æå¹æãããããã§ãã
- åçã®å¹³ååæ°ã«ã¤ãã¦ãã©ã¡ããå
±éãã¦å¢å ãã¦ãããã®ã®ãChat Vectorçã®ã»ããããé¡èã«å¢ãã¦ãã¾ãã
- ä½æçã«ãChat Vectorçã®åçã¯ç·ãã¦ä¸å¯§ã§ãMistral 7B v0.2 Instructã®å½±é¿ãæãããã¾ãã
- å®éã®åçä¾ãè¦ã¦ã¿ã¾ãã
- 以ä¸ã®ããã«ãå ã®ãChatNTQ 7Bãã§ã¯è§£ããªãã£ãè¨ç®åé¡ããã©ã³ãã ãã¼ã¸çã§ããChat Vectorçã§ãåãããã«æ£çã§ããããã«ãªãã¾ããã
ElyzaTasks100 è¨å26 |
質å: 太éããã¯ãããã5ã¤æã£ã¦ãã¾ããå½¼ã¯ããã«2ã¤ã®ãããã®ç®±ãè²·ãã¾ããã1ã¤ã®ç®±ã«ã¯3ã¤ã®ããããå ¥ã£ã¦ãã¾ãã太éããã¯ä½åã®ããããæã£ã¦ãã¾ããï¼ |
ChatNTQ-7B åçï¼ä¸æ£è§£Ãï¼ |
太éããã¯æåã«5ã¤ã®ããããæã£ã¦ãã¾ããããã®å¾ã2ã¤ã®ç®±ãè²·ããããããã«3ã¤ã®ããããå ¥ã£ã¦ããããã太éããã¯åè¨ã§8ã¤ã®ããããæã¤ãã¨ã«ãªãã¾ãããããã£ã¦ãå½¼ãæåã«æã£ã¦ãã5ã¤ã®ãããã¨ãå¾ããè²·ã£ã2ã¤ã®ç®±ã«å ¥ã£ã¦ãã3ã¤ã®ããããåãããã¨ãåè¨ã§8ã¤ã®ããããæã£ã¦ãããã¨ã«ãªãã¾ãã |
ChatNTQ 2x7B Random MoE åçï¼æ£è§£ãï¼ |
太éãããæåã«æã£ã¦ãã5ã¤ã®ãããã¨ã2ã¤ã®ç®±ã®åè¨åæ°ãè¨ç®ããå¿
è¦ãããã¾ãã 太éãããæåã«æã£ã¦ãã5ã¤ã®ãããã®æ°ã¯ã5ã§ãã 1ã¤ã®ç®±ã«ã¯3ã¤ã®ããããå ¥ã£ã¦ããããããã1ã¤ã®ç®±ã«ãåãæ°ã®3ã¤ã®ããããå ¥ã£ã¦ããã¨ä»®å®ããã¨ã2ã¤ã®ç®±ã«ã¯åè¨6ã¤ã®ããããå ¥ã£ã¦ãã¾ãã ãããã£ã¦ã太éãããæåã«æã£ã¦ãã5ã¤ã®ãããã¨ã2ã¤ã®ç®±ã«å ¥ã£ã¦ãã6ã¤ã®ããããåè¨ããã¨ã11åã®ããããæã£ã¦ãããã¨ã«ãªãã¾ãã ãããã£ã¦ã太éããã¯11åã®ããããæã£ã¦ãã¾ãã |
ChatNTQ-7B + Chat Vector åçï¼æ£è§£ãï¼ |
æåã«å¤ªéãããæã£ã¦ãããããã¯5ã¤ã§ããããã¦ã2ã¤ã®ç®±ãè²·ã£ã¦ããããã1ã¤ã®ç®±ã«3ã¤ãã¤å
¥ã£ã¦ããã®ã§ãåè¨ã§2 à 3 = 6ã¤ã®ãããã追å ããã¾ãã æåã®5ã¤ã¨åããã¦ã太éããã¯11åã®ããããæã£ã¦ãã¾ãã |
- MoEãã¼ã¸ã¨Chat Vectorã¨ããå ¨ãç°ãªãå¦çã§ãããå ¨è¬çãªæ§è½åä¸ã¯ã©ã¡ãã®ææ³ã«ããã¦ãæå¾ ã§ããããªå°è±¡ã§ãã
- ã©ã³ãã MoEãã¼ã¸ã®å ´åãããããç·ãã©ã¡ã¼ã¿æ°ã®å¢å ãæ§è½ã«å¯ä¸ãã¦ãã¦ãChat Vectorã®å ´åã¯è±èªã¢ãã«ã®è¿½å å¦ç¿ã¹ãã«ãæ§è½ã«å¯ä¸ãã¦ããã®ã ã¨æãã¾ãã
Chat Vectorã¢ãã«ã®ã©ã³ãã MoEãã¼ã¸
- ã¤ã¥ãã¦ãChat VectorãæãããChatNTQ 7Bããåç´ã«ã©ã³ãã MoEãã¼ã¸ããã¨ã©ããªãã§ãããããçµæã¯ä»¥ä¸ã®ã¨ããã§ãã
å称 | LightChat Assistant 2x7B |
ChatNTQ 7B + ChatVector Random MoE |
ChatNTQ 7B + ChatVector |
ChatNTQ 7B |
ãã©ã¡ã¼ã¿ | 2x7B | 2x7B | 7B | 7B |
éååã¿ã¤ã | Q8_0 | Q8_0 | Q8_0 | Q8_0 |
ElyzaTasks100 ã¹ã³ã¢å¹³å |
3.31 | 3.33 | 3.23 | 3.06 |
æ¨æºèª¤å·® | 0.14 | 0.15 | 0.15 | 0.15 |
å¹³ååæ° | 273.61 | 286.73 | 280.04 | 240.46 |
- ãã³ããã¼ã¯æ§è½ã§ã¯ãLightChatAssistant 2x7Bãã¨ã»ã¼åãã¬ãã«ã®ã¢ãã«ãå¾ããã¾ããã
- ãããã£ã¦ããLightChatAssistant 2x7Bãã®é«ãæ¥æ¬èªãã£ããæ§è½ããèªä½ã¯ããChat Vectorå¹æã +ããã¼ã¸ã«ãããã©ã¡ã¼ã¿å¢å å¹æãã®2ã¤ã ãã§èª¬æãããã¨ãã§ãããã§ãã
- ã¨ã¯ããå®éã®åºåãè¦æ¯ã¹ã¦ã¿ãã¨ãå°èª¬çæã¢ãã«ãæãåããããLightChatAssistant 2x7Bãã®åºåã®æ¹ããåä¸ã¢ãã«ãã¼ã¸ã®ãChatNTQ-ChatVector 2x7Bãããæããã«ã¯ãªã¨ã¤ãã£ãã§é¢ç½ãã¨æãã¾ãã
- ç°ãªãç¹åã¢ãã«ãæãåããããã¨ã®å¹æã¯ããã³ããã¼ã¯ã§æ¸¬å®ããããªãé¨åã«ããç¾ããã®ããããã¾ããã
å¥ã®ã¢ãã«ããChat Vectorãæ½åºãã¦ã¿ã
- ã¤ãã§ã«ãç°ãªãã¢ãã«ããChat Vectorãæ½åºãããã¨ã試ãã¦ã¿ã¾ãã
- å è«æããLightChatAssistant 2x7Bãã§å·®åæ½åºã«ä½¿ããããMistral 7B v0.2 Instructãã¯ãMistralå ¬å¼ã®ãã¡ã¤ã³ãã¥ã¼ã³ã¢ãã«ã§ãããä½æçã«ãåªãããã£ããã¢ãã«ã§ãã
- ãã ããã³ããã¼ã¯ä¸ã¯ãMistral 7B v0.2 Instructããä¸åãMistralç³»ã¢ãã«ãããã¤ãä½åºããã¦ãã¦ããã®ä»£è¡¨ä¾ãã¹ã¿ã³ãã©ã¼ã大å¦ç³»ã®ç 究è ãå ¬éãããStarling-LM-7B-betaãã§ãã
- Starling-7Bã·ãªã¼ãºã«ã¤ãã¦ã¯ã以ä¸ã®è¨äºã§æ¦è¦ã確èªãã¦ãã¾ãã
- ãStarling-LM-7B-betaãã¯ç¾å¨ããLMSYS Chatbot Arena Leaderboardãã§7Bã¢ãã«æé«ã¹ã³ã¢ã示ãã¦ãã¦ãè±èªãã£ããæ§è½ã§ã¯ãã®ã¯ã©ã¹ã®SOTAã§ãã
- 7Bã®ãªã¼ãã³ã¦ã§ã¤ãã人æã«ããè©ä¾¡ã§ GPT-3.5-Turboã»Claude 2.1ã»Gemini Pro ãªã©ä¸»è¦å社ã®åç¨ã¢ãã«ã¨åçã®è©ä¾¡ãå¾ãããã¨ããã®ã¯å°ã é©ãã§ãã
- ãã®ã¢ãã«ããChat Vectorãæ½åºãã¦æãåãããã°ãããæãã®æ¥æ¬èªãã£ããã¢ãã«ãåºæ¥ä¸ããäºæããã¾ãï¼ç´ 人ã®çºæ³ï¼ã
- ã¨ããããã§ãStarling-LM-7B-betaãã®Chat VectorããChatNTQ 7Bãã«æããã¢ãã«ãä½ãããã³ããã¼ã¯ã«ããã¦ã¿ã¾ããã
- ãªãããStarling-LM-7B-betaãã®Chat Vectorã¯1.0ã§æãã¦ãã¾ã£ããè±èªãæ··ãããªãã£ãã®ã§ã0.8ã§ã¯ãªã1.0ã®å¼·åº¦ãæ¡ç¨ãã¦ãã¾ãã
- çµæã¯ä»¥ä¸ã®ã¨ããã§ãã
å称 | Command R V01 |
LightChat Assistant 2x7B |
ChatNTQ + Staling ChatV |
ChatNTQ 7B |
ãã©ã¡ã¼ã¿ | 35B | 2x7B | 7B | 7B |
éååã¿ã¤ã | Q8_0 | Q8_0 | Q8_0 | Q8_0 |
ElyzaTasks100 ã¹ã³ã¢å¹³å |
3.42 | 3.31 | 3.42 | 3.06 |
æ¨æºèª¤å·® | 0.15 | 0.14 | 0.15 | 0.15 |
å¹³ååæ° | 194.02 | 273.61 | 340.82 | 240.46 |
- ã¹ã³ã¢ä¸ã¯ãLightChatAssistant 2x7Bããä¸åãã8bitéååãã¼ã¹ã§ãCommand R 35Bãã«ä¸¦ã¶ã¢ãã«ãåºæ¥ä¸ããã¾ããã
- åºååæ°ãå¤ãï¼å¹³å341åï¼ãã¨ã«ããä¸å¯§ãªãã£ãã対å¿ãã§ããããã«ãªã£ã¦ãã¾ããElyzaTasks100ã§ã¯åçã®ãæç¨æ§ããè©ä¾¡åºæºã«å«ã¾ããã®ã§ãè©ä¾¡è ã®GPT-4ããã®è¦³ç¹ããé«ãè©ä¾¡ãã¦ããããã§ãã
- ã¨ã¯ãããCommand R 35Bãã¯é¥ãã«ç°¡æ½ãªåçï¼å¹³å194åï¼ã§åãã ãã®ã¹ã³ã¢ãåºãã¦ããã®ã§ãç´ç²ãªãè³¢ããã«é¢ãã¦è¨ãã°æããã«ãCommand R 35Bããä¸ã ã¨æãã¾ãã
Starling LMã®Chat Vectorã¢ãã«ã®MoEãã¼ã¸
- ã§ã¯ããã®ã¢ãã«ãããã«MoEãã¼ã¸ããã°ãã£ã¨ã¹ã³ã¢ã¯ä¸ããã®ããã¨å®æã«èãã¦ãã£ã¦ã¿ã¾ããããçµæã¯ä»¥ä¸ã®ã¨ããã§ããã
å称 | LightChat Assistant 2x7B |
ChatNTQ + Staling ChatV Random MoE |
ChatNTQ + Staling ChatV |
ChatNTQ 7B |
ãã©ã¡ã¼ã¿ | 2x7B | 2x7B | 7B | 7B |
éååã¿ã¤ã | Q8_0 | Q8_0 | Q8_0 | Q8_0 |
ElyzaTasks100 ã¹ã³ã¢å¹³å |
3.31 | 3.18 | 3.42 | 3.06 |
æ¨æºèª¤å·® | 0.14 | 0.15 | 0.15 | 0.15 |
å¹³ååæ° | 273.61 | 351.42 | 340.82 | 240.46 |
- åºååæ°ããå¤å°å¢ãã¦ãããã®ã®ãGPT-4ã®è©ä¾¡ã¯éã«ä½ä¸ãã¦ãã¾ãã
- ä½æçã«ããå£ã®éãã«æèã追ãä»ãã¦ããªãããããªãã¡ãã¯ããªå°è±¡ãåããã¢ãã«ã«ãªã£ã¦ãã¾ãã¾ããã
- åä¸ã¢ãã«ã«å¯¾ãã¦æé©ãªChat Vectoræ¯çã¨MoEãã¼ã¸å¾ã«æé©ã«ãªãChat Vectoræ¯çãç°ãªãã¨ãã話ï¼ãã¡ãã®Aratakoããã®ã¢ãã«ã«ã¼ãåç §ï¼ããããå¿ ããã足ãç®ã«ã¯ãªããªãã®ã ã¨å確èªãã¾ããã
éæ
- ã¨ãããã¨ã§ãã²ã¨ã¾ããLightChatAssistant 2x7Bãã®å¾è¿½ããããªããMoEã¨ChatVectorã§ããã¤ãã®è©¦è¡é¯èª¤ããã¦ã¿ã¾ããã
- ãªããStarling ChatV 7Bã¢ãã«ãã«ã¤ãã¦ã¯ãã³ããã¼ã¯æ§è½ããã¾ãã«é«ããä½ãã®åèã«ããªãæ°ãããã®ã§ãHuggingFaceã«ã¢ãããã¦ã¿ã¾ããã
- Llama-3ã®å°åã¢ãã«ã®ãªãªã¼ã¹ãéè¿ã«è¿«ã£ã¦ããã¨ãã話ãããããã¡ãã楽ãã¿ã«ããªãããCVã¨MoEã§ä½ãã§ããããå¼ãç¶ã試ãã¦ããããã¨æãã¾ãã
Â