2024-01-01ãã1å¹´éã®è¨äºä¸è¦§
è¨èªã¢ãã«ã®ãè¨æ¶ãã管çããæ¹æ³ã¨ãã¦ã¯ãï¼ï¼è¿½å å¦ç¿ãç¥èç·¨éã«ãã£ã¦LLMèªä½ã®ç¥èæ´æ°ã試ã¿ãæ¹æ³ã¨ãï¼ï¼å¤é¨ã®è¨æ¶ãã¼ã¿ããå¿ è¦æ å ±ãé©ææ¤ç´¢ãã¦ã¢ãã«ã«æ¸¡ãæ¹æ³ãã®2ã¤ã®æ¹åæ§ãããã¾ãã ãã®è«æã¯å¾è ã®ç³»çµ±ã«å±ããç 究ã§ã人éã®â¦
ãç´æ¥é¸å¥½æé©åï¼DPOï¼ãã¯ããããã¢ã©ã¤ã³ã¡ã³ãã®ç®çã§ä½¿ããã¦ããLLMã®å¦ç¿ææ³ã§ãã åä¸ã®æ示ã«å¯¾ããæçãªåçä¾ã¨æ害ãªåçä¾ã®ä¸¡æ¹ãæ示ãããã¨ã§ãã¢ãã«ãéçºè ã«ã¨ã£ã¦å¥½ã¾ããæåãã¨ãããå¹ççã«èª¿æ´ãã¾ãã ä»ã®ã¢ã©ã¤ã³ã¡ã³â¦
å°ãå¤ãã§ããããã¡ã¤ã³ãã¥ã¼ãã³ã°ã®çµçï¼The End of Fine-tuningï¼ãã¨ããè¨äºãé¢ç½ãã£ãã®ã§ãç°¡åãªã¡ã¢ã§ãã www.latent.space ãã®è¨äºã«ç»å ´ããFast.aiã®Jeremy Howardããã¯ãäºåå¦ç¿â追å å¦ç¿âRLHFãã®ããã«ããã§ã¼ãºãã¨ã«ãã¼ã¿ã»ãâ¦
å æ¥ãLlama-3-70Bãåãè¾¼ã¿ï¼Pruningï¼ã§42Bã«ç¸®ããã¢ãã«ãç»å ´ãã¨ããRedditã®æ稿ã話é¡ã«ãªã£ã¦ãã¾ããã æ稿è ã¯ããªãã¿ã®kindacognizantï¼kalomazeï¼ããã§ãããã¢ãã«ä½æè ã¯å¥ã®æ¹ã®ããã§ããã¢ãã«ã®HuggFaceã®Repoã¯ãã¡ãã huggingfacâ¦
éååæã®ã¢ãã«å£åãæå¶ããéè¦åº¦è¡åï¼iMatrix; Importance Matrixï¼è¨ç®ã®è©±é¡ã§ãã æè¿ã¯HuggingFaceã«ã¢ãããããGGUFãå¤ããiMatrixçã¨ãªã£ã¦ãã¾ããããããã®éååã§ãã使ããã¦ããiMatrixè¨ç®ç¨ãã¼ã¿ã»ããã¯ä»¥ä¸ã®2種é¡ã®ããã§ããâ¦
ãMerggooãã¨ããLLMãã¼ã¸ç¨ã®æ°ããã©ã¤ãã©ãªãã§ããããã§ãéçºè ã®æ¹ãRedditã§ãPRãã¹ããä¸ãã¦ãããã¾ãã ãããããã¨Megekitã«ä»£ãããã®ã«ãªãã®ããããã¾ãããMoEã®ã«ã¼ãã£ã³ã°ã®å¦ç¿ããLoRA Adapterã®æ··åãªã©ã é¢ç½ãããªã®ã§å¾ã§â¦
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 sc-baâ¦
ååChat Vectorã«ã¤ãã¦ç°¡åã«äºç¿ããã®ã§ãã¨ãããããLightChatAssistant 2x7Bãã®ä½ææé ãåç¾ãã¦ã¿ããã¨æãã¾ãã ä½è ãããã¢ãã«ã«ã¼ãã§ä¸å¯§ã«èª¬æãã¦ãã ãã£ã¦ããã®ã§ãåºæ¬çã«ããããªããã ãã§ããã¾ãã¯ã¾ã£ããåãã¢ãã«ãä½ã£ã¦â¦
ååã®è¨äºã«ã¤ã¥ãããLightChatAssistant 2x7Bãã«é«ããã£ããæ§è½ããããããè¦ç´ ã«ã¤ãã¦èª¿ã¹ã¦ãã¾ãã åºæ¬çã«ã¯ã1ï¼ãã¼ã¹ã¢ãã«ã¨ãã¦ã®ãChatNTQ JA 7Bãã®æ¥æ¬èªæ§è½ã®é«ãã2) MoEãã¼ã¸ã§ç·ãã©ã¡ã¼ã¿æ°ãå¢ãããã¨ã«ããå ¨è¬çãªæ§è½åä¸â¦
ååã®è¨äºã§ããLightChatAssistant 2x7Bããæ¥æ¬èªãã£ããã¢ãã«ã®ãã³ããã¼ã¯ã§ãããªãé«ãã¹ã³ã¢ã示ããã¨ã確èªãã¾ããï¼ã¹ã³ã¢ä¸ã¯Cohereã®ãCommand-R 35Bãã«è¿ãæ°´æºã§ãï¼ã ä½æçã«ããã³ããã¼ã¯çã«ãåªããæ¥æ¬èªãã£ããã¢ãã«ã§ããµã¤â¦
ååã®è¨äºã§è©¦ããæ¥æ¬èªãã£ããã¢ãã«ãChatNTQ-JA-7B-v0.1ãã¨ããã®MoEã¢ãã«ãLightChatAssistant 2x7Bï¼æ¹ç§°ããï¼ãã«ã¤ãã¦ãããªãæ§è½ãè¯ããããªæ触ãå¾ãããã®ã§ã追å ã§ãã¹ããã¦ã¿ã¾ããã LLMã®æ¥æ¬èªãã£ããæ§è½ã測ããã³ããã¼ã¯ã¨â¦
ãchatntq_chatvector-MoE-Antler_chatvector-2x7Bchatntq_chatvector-MoE-Antler_chatvector-2x7Bãã¨ããåªæã®ãããªæ¥æ¬èªMoEã¢ãã«ã話é¡ã«ãªã£ã¦ãã¾ããã https://t.co/tmcIFgrObQ2x7Bã®æ¥æ¬èªãã£ããã»ããã«å°ç¨é«æ§è½ã¢ãã«ãAntler-7Bã¨chatntqâ¦
ä»ããç¥ã£ãã®ã§ããã2é±éã»ã©åã«llama.cppã§ã¢ãã«ãGPUã«é¨åãªããã¼ãããå ´åã®ããã³ããå¦çé度ãå¤§å¹ ã«åä¸ããã¦ãã¾ããã github.com å¾æ¥ã®llama.cppã§ã¯GPUãã«ãªããã¼ãããé¨åãªããã¼ãã«ç§»è¡ããã¨æ¥æ¿ã«ããã³ããå¦çï¼PPï¼ãé ãâ¦
以åã«ãåãä¸ãã¦ãã件ã§ãããç¾å¨ã®llama.cppã§ã¯éè¦åº¦è¡å(Importance Matrix)è¨ç®ãå©ç¨ãããã¨ã§éåå精度ãæ¹åã§ãã¾ãã ç¹ã«4bit以ä¸ã®ä½bitéååãè¡ãå ´åã¯ããã®iMatrixçã®éååãæ¨å¥¨ããã¾ãï¼Metalç°å¢ãªã©ã§ã¯æ¨è«é度ãé ããªãâ¦
æè¿ãèªä½ã®æ示å¿çãã¼ã¿ã»ããã使ã£ã微調æ´ã«ãã£ã¦LLMã«ãã¼ã½ãã©ã¤ãºãããç¥èã追å ãããã¨ã試ã¿ã¦ããã®ã§ããããã®éã«ã¢ãã«ã«æ¤ãä»ããããéå°ãªã¢ã©ã¤ã³ã¡ã³ããç¥è追å ã®éå£ã«ãªãå ´åãããã¾ãã ä¾ãã°ã¢ãã«ã«å¯¾ããUSER: 好ãâ¦
ã馴æã¿ã®LMSYS Chatbot Arena ELOã©ã³ãã³ã°ãæ´æ°ããã¦ãã¾ããã [Arena Update]70K+ new Arena votesï¸ are in!Claude-3 Haiku has impressed all, even reaching GPT-4 level by our user preference! Its speed, capabilities & context length are uâ¦
ChatGPTã®ãããªæ¶è²»è åãAIãã£ãããµã¼ãã¹ã®åçååé¡ã«é¢ããè¨äºãRedditã§å ±æããã¦ãã¾ããã www.businessinsider.com è¨äºã«ãã㨠æè¿ãInflection AIãã¨ããAIã¹ã¿ã¼ãã¢ãããã主è¦ã¡ã³ãã¼ãMicrosoftã«å¼ãæãããä¼ç¤¾ãç¦è§£ãããã£ã¦â¦
ããã¡ã¤ã³ãã¥ã¼ã³ç¨ã®ãã¼ã¿ã«ã¯äºåå¦ç¿ç¨ãã¼ã¿ãããã¯ã¹ãã¹ããã¨ãããã¹ããredditã«ä¸ãã£ã¦ãã¾ãããæ稿è ã¯kobold.cppãªã©ã®éçºã«ãé¢ãã£ã¦ããkindacognizant (kalomaze)ããã§ãã æ稿ã®è¦æ¨ è¨èªã¢ãã«ã®ãã¡ã¤ã³ãã¥ã¼ã³ã¯åºæ¬çã«ãå â¦
www.youtube.com Mistral AI 㨠Figma ã®CEOã®å¯¾è«ã«é¢ããæ稿ãRedditã«ä¸ãã£ã¦ããï¼æåèµ·ããã®ãªã³ã¯ãè²¼ããã¦ããï¼ãç®ãéãã¦æ°ã«ãªã£ãç¹ãé©å½ã«ã¡ã¢ãã¦ããã Llama-7Bã®ãããªå°åã®ã¢ãã«ã¯ã³ãã¥ããã£ã®éè¦ã大ããä¸æ¹ã§ãæ¹åã®ä½å°â¦
LoRAãã¡ã¤ã³ãã¥ã¼ã³ã§ã¯æ§ã ãªãã¤ãã¼ãã©ã¡ã¼ã¿ãããã¾ããã¢ãã«ã¨ãã¼ã¿ã»ããã«åã£ããã©ã¡ã¼ã¿ãé¸ã¶ãã¨ã§ãå¦ç¿é度ã»ç²¾åº¦ãå¤ããã¾ãã ä»æ¥ã¯ä¸»è¦ãªãã¤ãã¼ãã©ã¡ã¼ã¿ã®ä¸ã¤ã§ããLoRAã©ã³ã¯ ï¼rï¼ãæ°ã«ãªã£ãã®ã§ãç°¡åãªåå¿é²ãæ¸ãã¦ãâ¦
ããæ°æ¥ãStable Knowledge Editingããåèã«ããªãããLoRAãã¡ã¤ã³ãã¥ã¼ã³ã«ããLLMã¸ã®ç¥èã®è¿½å ã試ãã¦ãã¾ãã LoRAã®ãã¤ãã¼ãã©ã¡ã¼ã¿èª¿æ´ã®ã³ãã調ã¹ããªãã§ããDoRAï¼éã¿å解LoRAï¼ãã¨ããå¥ã®LoRAæ´¾çææ³ã®åå¨ãç¥ãã¾ãããHuggingFaâ¦
2024å¹´2æã«æ稿ãããä¸å½ç§å¦é¢å¤§å¦ã®ç 究è ã«ããarXivè«æã§ãã ãã¡ã¤ã³ãã¥ã¼ã³ã»ãã¼ã¹ã®ç¥èç·¨éææ³ã§ãããStable Knowledge Editingããææ¡ããããã§ãæ¢åã®ç¥èç·¨éææ³ã¨æ¯ã¹ãå ´åã®æç¨æ§ã主張ãã¦ãã¾ãã arxiv.org æ¦è¦ 大è¦æ¨¡è¨èªã¢â¦
1. GGUFï¼å ¬å¼docï¼ github.com 2. ãã©ãããã©ã¼ã å¥GGUF対å¿è¡¨ github.com 3. ã©ã®GGUFãé¸ã¹ã°ããã§ããï¼ï¼åæ²ï¼ GGUF quantizations overview · GitHub
github.com ã«ããã®AIã¹ã¿ã¼ãã¢ããCohereãå æ¥å ¬éããå¤è¨èªLLMã®Command-Rããææ°ã®llama.cppã§ãµãã¼ãããã¾ããã éçºããCohereã¯LLMã¹ã¿ã¼ãã¢ããã¨ãã¦ã¯ããã¨æåã©ããã§ãããOpenAI/Anthropic/Mistralã®å é éå£ã«ã¯å¾ããåã£ã¦ããæâ¦
LLMã®ç¥èç·¨éï¼Knowledge Editingï¼ã®ããã¾ãããã£ããææ¡ãããããé©å½ãªãµã¼ãã¤è«æã«ç®ãéãã¦ã¿ããã¨æãã¾ãã arXivã«ä¸ãã£ã¦ãã2023å¹´10æã®ãã¼ã¸ãã¢å¤§å¦ã®ç 究è ã«ããè«æã大è¦æ¨¡è¨èªã¢ãã«ã®ç¥èç·¨éã«é¢ãããµã¼ãã¤ï¼Knowledge Ediâ¦
å æ¥ãAnthropic ãçºè¡¨ããã¯ãã¼ãºãã®å¤§è¦æ¨¡è¨èªã¢ãã« Claude 3 (Opus) ããæ¬å½ã«GPT-4 è¶ ãã®æ§è½ãããã¨è©±é¡ã«ãªã£ã¦ãã¾ãã Chatbot Arena Leaderboard ã®ç´è¿éè¨ã§ã¯ãåé ã®è¡¨ã®ã¨ãã GPT-4 Turbo ã®å¾å¡µãæãã¦ãããã®ã®ãæè»ã§ç©æ¥µçãªå¯¾â¦
IBMããibm/merlinite-7bãã¨ããMistral 7Bãã¼ã¹ã®ãã¡ã¤ã³ãã¥ã¼ã³ã¢ãã«ãå ¬éããåæã«ãã®ãã¡ã¤ã³ãã¥ã¼ã³ææ³ã«é¢ããarXivè«æãæ稿ãã¦ãã¾ãï¼ãLAB: ãã£ãããããã®ããã®å¤§è¦æ¨¡ã¢ã©ã¤ã¡ã³ããï¼ã arxiv.org æ¦è¦ æ¬ç 究ã§ã¯ã大è¦æ¨¡è¨èªã¢â¦
https://github.com/ggerganov/llama.cpp/pull/5747 llama.cpp ã§ã¯æè¿ãikawrakowæ°ã«ããéååææ³ã®ã¢ãããã¼ããç±å¿ã«è¡ããã¦ãã¾ãã æ°ããéååã®å®è£ ãéãªãå人çã«åããã¥ãããªã£ã¦ããã®ã§ãç°¡åã«æ´çãã¦ãããã¨æãã¾ãã quantize.â¦
ããã¡ã¤ã³ãã¥ã¼ãã³ã°ãæ¤ç´¢(Retrieval)ãï¼LLMã«ãããç¥è追å ææ³ã®æ¯è¼ãã¨ããarXivè«æãä¸ãã£ã¦ãã¾ãããå·çè ã¯Microsoft Israelã®ç 究ã°ã«ã¼ãã§ãã arxiv.org å æ¥ãMicrosoft ã®å¥ã®ã°ã«ã¼ãããRAG vs ãã¡ã¤ã³ãã¥ã¼ãã³ã°ï¼ãã¤ãã©ã¤ã³â¦
å¹´æãã«llama.cppã«å®è£ ããããImportance Matrixï¼éè¦åº¦è¡åï¼ï¼ãã使ç¨ããgguféååã«ã¤ãã¦èª¿ã¹ã¦ã¿ã¾ããã Importance Matrixã¯ãllama.cppã®ikawrakowæ°ãåãçµãã§ããä¸é£ã®éåå精度æ¹åã®ä¸å¿çãªã¢ã¤ãã£ã¢ã®ããã§ãç¹ã«2-3bitã®æ¥µç«¯ãªâ¦