
Orca : GPT-4ã®æšè«éçšãæš¡å£ããChatGPTã«å¹æµããæ§èœãæãã13Bã®OSSèšèªã¢ãã«
Orca: Progressive Learning from Complex Explanation Traces of GPT-4ãšããMicrosoft Researchãçºè¡šããè«æã®äžéšã玹ä»ããŸãã
åŸæ¥ã®instruction-tuned model (Vicuna-13Bãªã©)ãšã¯ç°ãªããGPT-4ã®ãããªå€§èŠæš¡èšèªã¢ãã«ã®ã¹ããããã€ã¹ãããã®æšè«éçšãåŠç¿ãããããšã§ã軜éãªã¢ãã«ãªããChatGPT (3.5)ã«å¹æµãã粟床ãå®çŸããã¢ãã«ãšã®ããšã§ãã
èŠçŽ( by ChatGPT)
Microsoft Researchãéçºãã13åãã©ã¡ãŒã¿ã®ã¢ãã«Orcaã¯ãGPT-4ããã®è±å¯ãªä¿¡å·ã説æã®ãã¬ãŒã¹ãã¹ããããã€ã¹ãããã®æèéçšãå©çšããŠãå€§èŠæš¡åºç€ã¢ãã«ïŒLFMsïŒã®æšè«éçšãæš¡å£ããããã«åŠç¿ããŸããOrcaã¯åŸæ¥ã®æå 端ã®instruction-tuned model (Vicuna-13Bãªã©)ãäžåããSATãLSATãGREãGMATãªã©ã®å°éçãªåŠè¡è©Šéšã§ç«¶äºåã®ããããã©ãŒãã³ã¹ã瀺ããŸãããã®ç ç©¶ã¯ãã¹ããããã€ã¹ãããã®èª¬æããåŠç¿ããããšããã¢ãã«ã®èœåãšã¹ãã«ãæ¹åããããã®ææãªæ¹åæ§ã§ããããšã瀺ããŠããŸããã¢ãã«ã®éã¿ã®å·®åã¯ãä»åŸLLaMAã®ãªãªãŒã¹ããªã·ãŒã«åŸã£ãŠå ¬éãããŸãhttps://aka.ms/orca-lmã
æŠèŠ
æè¿ã®ç ç©¶ã§ã¯ãæš¡å£åŠç¿ãå©çšããŠå€§èŠæš¡ãªåºç€ã¢ãã«ïŒLFMsïŒãçæããåºåãåèã«ãå°ããªã¢ãã«ã®æ§èœãåäžãããããšã«çŠç¹ãåœãŠãããŠããŸãããããã®ã¢ãã«ã®å質ã«åœ±é¿ãåãŒãèŠå ã¯å€ããæµ
ãLFMã®åºåããåŸãããéå®çãªæš¡å£ä¿¡å·ãå°èŠæš¡ãªå質ã®ãã¬ãŒãã³ã°ããŒã¿ããããŠç¹ã«å³æ Œãªè©äŸ¡ã®æ¬ åŠã«ãããå°ããªã¢ãã«ãLFMsã®ã¹ã¿ã€ã«ãæš¡å£ããããšã¯ãã£ãŠãããã®æšè«ããã»ã¹ãåŠã¶ããšãå°ãªããããå°ããªã¢ãã«ã®èœåãé倧è©äŸ¡ãããããšãæããããŸãã
ãããã®èª²é¡ã«å¯ŸåŠããããã«ãæã
ã¯Orcaãšãã130åãã©ã¡ãŒã¿ã®ã¢ãã«ãéçºããLFMsã®æšè«ããã»ã¹ãæš¡å£ããããã«åŠç¿ããŸããOrcaã¯ãGPT-4ããã®è±å¯ãªä¿¡å·ãåŠã³ã説æã®ãã¬ãŒã¹ã段éçãªæèããã»ã¹ãããã³ChatGPTããã®æåž«ã®æ¯æŽã«ãã£ãŠæ¡å
ããããã®ä»ã®è€éãªæç€ºãå«ã¿ãŸãããã®æ®µéçãªåŠç¿ãä¿é²ããããã«ãæã
ã¯å€§èŠæš¡ã§å€æ§ãªæš¡å£ããŒã¿ã«ã¢ã¯ã»ã¹ããæ
éãªãµã³ããªã³ã°ãšéžæãè¡ããŸãã
Orcaã¯ãVicuna-13Bãªã©ã®åŸæ¥ã®æå
ç«¯ã®æç€ºèª¿æŽã¢ãã«ã倧å¹
ã«äžåããBigBench Hard (BBH)ãªã©ã®è€éãªãŒãã·ã§ããæšè«ãã³ãããŒã¯ã§100ïŒ
以äžãAGIEvalã§ã¯42ïŒ
ã®åäžã瀺ããŸããããã«ãOrcaã¯BBHãã³ãããŒã¯ã§ChatGPTãšåçã®æ§èœãéæããSATãLSATãGREãGMATãªã©ã®å°éçããã³åŠè¡çãªè©Šéšã§ç«¶åæ§èœã瀺ããŸãïŒæé©åãããã·ã¹ãã ã¡ãã»ãŒãžãšã®ã®ã£ããã¯4ãã€ã³ãïŒããã ããGPT-4ã«ã¯åã³ãŸãããç§ãã¡ã®ç ç©¶ã¯ã人éãŸãã¯ããé«åºŠãªAIã¢ãã«ã«ãã£ãŠçæãããæ®µéçãªèª¬æããåŠã¶ããšããã¢ãã«ã®èœåãšã¹ãã«ãåäžãããããã®ææãªæ¹åã§ããããšã瀺ããŠããŸãã
Introduction
13åãã©ã¡ãŒã¿ã®instruction-tuned modelã§ããVicuna [9]ïŒããŒã¹ãšããŠLLAMA-13B [10]ã䜿çšïŒã¯ãOpenLLM3ãChatArena4ã®ãªãŒããŒããŒãã§ã®ããã©ãŒãã³ã¹ãªã©ããããã®ãã¡ããªãŒã®äžã§æãåªããã¢ãã«ã®äžã€ãšåºãèªèãããŠããŸãã
å³1ã«ç€ºãããã«ãGPT-4ãå€æåºæºãšããŠäœ¿çšãããšããåºã䜿çšãããŠããè©äŸ¡æ¹æ³ã§ã¯ãVicunaã¯ChatGPTã®å質ã92%ä¿æããŠãããšç€ºåãããŠããŸãããããã人éã®ã©ãã«ã«å¯Ÿããæšè«ãã³ãããŒã¯ã§ããå³å¯ãªè©äŸ¡ãè¡ããšãVicunaã¯å°éçãªåŠè¡è©Šéšã§ChatGPTã®å質ã64%ããä¿æããŠããªãããšãããããŸãïŒå³2åç §ïŒããããŠBigBench-hard [11]ã®ãããªè€éãªãã³ãããŒã¯ã§ã¯ChatGPTã®å質ã48%ããä¿æããŠããŸããïŒå³3åç §ïŒ5ã
ãã®äžäžèŽã¯ãå°ããªLLMãšã®æ¢åã®è©äŸ¡ãããã³ã«ã®éçã匷調ããã ãã§ãªããæšè«ãšçè§£èœåã«ããããã®é¡èãªé ããæããã«ããŠããŸããæ¬è³ªçã«ããããã®ã¢ãã«ã¯éåŒãããããŸããããå¿ ãããå ç¢ãªæšè«ã¹ãã«ãæã£ãŠããããã§ã¯ãªããããããŸãããæ¬ç ç©¶ã§ã¯ããããã®ã®ã£ããã®èåŸã«ããããã€ãã®çç±ãè°è«ãããããã«å¯ŸåŠããããã®æŠç¥ãææ¡ããŸãã



åŸæ¥ã®æ¹æ³ã®åé¡ç¹
ã¿ã¹ã¯ã®å€æ§æ§ãéãããŠããïŒSelf-Instructããã»ã¹ã¯ãLFMã«æ°ããæç€ºãçæãããããã®åæã®ããã³ããã䜿çšããŸããããããçæãããã¯ãšãªã¯å€æ§æ§ãšè€éãã«å¶éãããå¯èœæ§ããããŸãã
ã¿ã¹ã¯ã®å€æ§æ§ãšããŒã¿ã®ã¹ã±ãŒãªã³ã°ïŒShareGPTã®äººéãå¯äžããäŒè©±ã¯è²ŽéãªããŒã¿æºã§ãããå¶éããããŸãããããã¯åµé çãªã³ã³ãã³ãçæãšæ å ±æ¢æ±ã®ã¯ãšãªãä»ã®ã¿ã¹ã¯ãããåªå ããåŸåããããŸãããããã£ãŠããã®ãããªèªç¶ãªäŒè©±ã§èšç·Žãããã¢ãã«ã¯ãã¹ã¿ã€ã«ãæããããšãã§ããŸãããLFMã®æšè«ããã»ã¹ã¯æããããŸããã
éå®çãªæš¡å£ä¿¡å·ïŒæ¢åã®æ¹æ³ã¯ãæåž«ã¢ãã«ã«ãã£ãŠçæãããâšã¯ãšãªãå¿çâ©ã®ãã¢ããåŠç¿ããæš¡å£åŠç¿ã«äŸåããŠããŸããããããããã¯æåž«ã®æšè«ããã»ã¹ã远跡ããããã®ä¿¡å·ãéå®çã§ããããšãæå³ããŸãã
è©äŸ¡ïŒå°èŠæš¡ã¢ãã«ã®LFMã«å¯Ÿããæç€ºãã¥ãŒãã³ã°ã®å è¡ç ç©¶ã¯ãè©äŸ¡ãããã³ã«ãå€§å¹ ã«å¶éãããŠããŸãããããã¯ãã°ãã°ãGPT-4ã«ããèªåè©äŸ¡ã«äŸåããŠããŸãããããããã®ã¢ãããŒãã«ã¯ããã€ãã®æ¬ ç¹ããããŸããããšãã°ããã¹ãã»ããã®ãµã€ãºãå°ããïŒVicunaã§ã¯80ã®æç€ºãWizardLMã§ã¯218ã®æç€ºïŒãGPT-4ã倿è ãšããŠã®ãã€ã¢ã¹ããããªã©ã§ãã
ãã®ç ç©¶ã®äž»èŠãªè²¢ç®ïŒ
説æãã¥ãŒãã³ã°ïŒç§ãã¡ã¯ãæåž«ãå¿çãçæããéã®æšè«ããã»ã¹ã説æããGPT-4ããã®è©³çްãªå¿çããâšã¯ãšãªãå¿çâ©ã®ãã¢ã«è¿œå ããŸããããã«ãããåŠç¿è ã«è¿œå ã®åŠç¿ä¿¡å·ãæäŸãããŸããç§ãã¡ã¯ãã·ã¹ãã æç€ºïŒäŸãã°ãã5æ³ã®åäŸã«èª¬æããããã«ããšãããã¹ããããã€ã¹ãããã§èããŠãããªãã®å¿çãæ£åœåããŠãã ããããªã©ïŒãå©çšããŠããã®ãããªèª¬æãåŒãåºããŸããããã¯ãéåžžã®æç€ºãã¥ãŒãã³ã°ãšã¯å¯Ÿç §çã§ãéåžžã®æç€ºãã¥ãŒãã³ã°ã¯ãããã³ãããšLFMã®å¿çã®ã¿ãåŠç¿ã«äœ¿çšããLFMã®ãæèãããã»ã¹ãæš¡å£ããæ©äŒãã»ãšãã©æäŸããŸããã
ã¿ã¹ã¯ãšæç€ºã®ã¹ã±ãŒãªã³ã°ïŒç§ãã¡ã¯ãFlan 2022 Collection [19]ãå©çšããŸããããã¯ãã¿ã¹ã¯ãšæç€ºã®åºç¯ãªå ¬éã³ã¬ã¯ã·ã§ã³ãæäŸããŸããç¹ã«ãç§ãã¡ã¯ãé«å質ã®ãã³ãã¬ãŒããé«åºŠãªãã©ãŒããããã¿ãŒã³ãããŒã¿æ¡åŒµãè£å®ããFLANv2ã䜿çšããŸããFLANã«ã¯æ°åäžã®æç€ºãå«ãŸããŠããŸãããç§ãã¡ã¯ãã¿ã¹ã¯ã³ã¬ã¯ã·ã§ã³ããéžæçã«ãµã³ããªã³ã°ããŠãã¿ã¹ã¯ã®å€æ§ãªããã¯ã¹ã圢æãããããããã«ãµããµã³ããªã³ã°ããŠè€éãªããã³ãããçæããŸãããããã®ããã³ããã¯ãChatGPTãGPT-4ã®ãããªLFMã«ã¯ãšãªãéãããã«äœ¿çšãããè±ãã§å€æ§ãªèšç·Žã»ãããäœæããŸããç§ãã¡ã¯500äžã®ChatGPTå¿çãåéãããã®ãã¡100äžãããã«ãµã³ããªã³ã°ããŠGPT-4ã®å¿çãååŸããŸããç§ãã¡ã¯ãChatGPTãæåž«ã¢ã·ã¹ã¿ã³ããšããŠã©ã®ããã«é²è¡åŠç¿ã«åœ¹ç«ã€ãã瀺ããŸãã
è©äŸ¡ïŒç§ãã¡ã¯ãOrcaã®çæèœåãæšè«èœåãçè§£èœåãã以äžã®ç¯å²ã®èšå®ã§è©äŸ¡ããŸãïŒ(i) VicunaãWizardLMããããŠçŽ æŽãããããã³ããã³ã¬ã¯ã·ã§ã³ããã®æ¢åã®è©äŸ¡ã»ããã§ã®GPT-4ã«ããèªåè©äŸ¡ã(ii) Big-Bench Hard [4]ãTruthfulQA [20]ãªã©ã®åŠè¡çãªãã³ãããŒã¯ã(iii) AGIEval [1]ããã®SATãLSATãGREãGMATãªã©ã®ãããã§ãã·ã§ãã«ããã³ã¢ã«ãããã¯ãªè©Šéšã(iv) ToxiGen [21]ãçšããå®å šæ§è©äŸ¡ã§ãç°ãªããã€ããªãã£ã°ã«ãŒãã«å¯Ÿããæå®³ãªèšèªçæãšãã€ãã¹ããŒãæ€åºããã¹ãããŸãã
æåŸã«ãOpenAI LFMsã®ChatGPTãšGPT-4ããããŠæç€ºèª¿æŽãããå°åã¢ãã«ã§ããVicunaãšæ¯èŒããŠãOrcaã®çæèœåãšæšè«èœåãã±ãŒã¹ã¹ã¿ãã£ã§ç€ºããŸãã