Microsoft¤¬¾®µ¬ÌϸÀ¸ì¥â¥Ç¥ë¡ÖPhi-2¡×¤ò¥ê¥ê¡¼¥¹¡¢¾®µ¬ÌϤʤΤ˺ÇÂç25ÇÜ¥µ¥¤¥º¤Î¥â¥Ç¥ë¤ÈƱÅù°Ê¾å¤ÎÀǽ
2023ǯ11·î¤Î¥¤¥Ù¥ó¥È¡ÖMicrosoft Ignite 2023¡×¤Çȯɽ¤µ¤ì¤¿¸À¸ì¥â¥Ç¥ë¤Î¡ÖPhi-2¡×¤¬¥ê¥ê¡¼¥¹¤µ¤ì¤Þ¤·¤¿¡£¥Ñ¥é¥á¡¼¥¿¡¼¿ô¤Ï27²¯¤Ç¾®µ¬ÌϤʥâ¥Ç¥ë¤È¤Ê¤Ã¤Æ¤¤¤ë¤â¤Î¤Î¡¢ºÇÂç¤Ç25ÇܤΥâ¥Ç¥ë¤ÈƱÅù¤ÎÀǽ¤òȯ´ø¤Ç¤¤Þ¤¹¡£
Phi-2: The surprising power of small language models - Microsoft Research
Today, we share our teams¡Ç latest contributions, Phi-2 and promptbase.
Phi-2 outperforms other existing small language models, yet it¡Çs small enough to run on a laptop or mobile device. https://t.co/wLhUeRsByL— Microsoft Research (@MSFTResearch) 2023ǯ12·î12Æü
¡ÖPhi¡×¤ÏMicrosoft Research¤Îµ¡³£³Ø½¬´ðÈ×¥Á¡¼¥à¤¬³«È¯¤·¤Æ¤¤¤ëTransformer¥Ù¡¼¥¹¤Î¾®µ¬ÌϸÀ¸ì¥â¥Ç¥ë¥·¥ê¡¼¥º¤Ç¤¹¡£ºÇ½é¤Î¥â¥Ç¥ë¤Ç¤¢¤ë¡ÖPhi-1¡×¤Ï13²¯¥Ñ¥é¥á¡¼¥¿¡¼¤Ç¡¢´û¸¤Î¾®µ¬ÌϸÀ¸ì¥â¥Ç¥ë¤ÎÃæ¤ÇPython¥³¡¼¥Ç¥£¥ó¥°¤Ë¤ª¤¤¤ÆºÇÀèü¤Î¥Ñ¥Õ¥©¡¼¥Þ¥ó¥¹¤òãÀ®¤·¤Þ¤·¤¿¡£Phi-1¤òµ¯ÅÀ¤Ë°ìÈÌŪ¤Ê¿äÏÀ¤È¸À¸ìÍý²ò¤ÎǽÎϤò¸þ¾å¤µ¤»¤¿¥â¥Ç¥ë¤¬¡ÖPhi-1.5¡×¤Ç¡¢¥Ñ¥é¥á¡¼¥¿¡¼¿ô¤¬13²¯¤È¾®¤µ¤¤¤Ê¤¬¤é¤â5ÇÜÂ礤¤¥â¥Ç¥ë¤ÈƱÅù¤Î¥Ñ¥Õ¥©¡¼¥Þ¥ó¥¹¤òȯ´ø¤·¤Æ¤¤¤Þ¤¹¡£
º£²ó¥ê¥ê¡¼¥¹¤µ¤ì¤¿Phi-2¤Ï27²¯¥Ñ¥é¥á¡¼¥¿¡¼¤Î¥â¥Ç¥ë¤Ç¡¢¥Ñ¥é¥á¡¼¥¿¡¼¿ô¤¬130²¯Ì¤Ëþ¤Î´ðËܸÀ¸ì¥â¥Ç¥ë¤ÎÃæ¤ÇºÇÀèü¤Î¥Ñ¥Õ¥©¡¼¥Þ¥ó¥¹¤òãÀ®¤Ç¤¤¿¤È¤Î¤³¤È¡£¤½¤Î¤Û¤«¡¢¤µ¤Þ¤¶¤Þ¤Ê¥Ù¥ó¥Á¥Þ¡¼¥¯¤ò·×¬¤¹¤ë¤ÈºÇÂç¤Ç¥µ¥¤¥º¤¬25ÇÜÂ礤¤¥â¥Ç¥ë¤ÈƱÅù°Ê¾å¤ÎÀǽ¤òȯ´ø¤·¤Þ¤·¤¿¡£¥â¥Ç¥ë¤Î¥µ¥¤¥º¤¬¾®¤µ¤¤¤¿¤á¥È¥ì¡¼¥Ë¥ó¥°¤ä¿äÏÀ¤Ê¤É¤Î¥³¥¹¥È¤òÍÞ¤¨¤ë¤³¤È¤¬¤Ç¤¡¢¸À¸ì¥â¥Ç¥ë¤ò¸¦µæ¤¹¤ë¤Î¤ËŬ¤·¤Æ¤¤¤ë¤È¤Î¤³¤È¡£
¾®¤µ¤¤¥â¥Ç¥ë¤ÇÂ礤¤¥â¥Ç¥ë¤ËɤŨ¤¹¤ëÀǽ¤ò½Ð¤¹¤¿¤á¤Ë¡¢Microsoft¤Ï¥È¥ì¡¼¥Ë¥ó¥°¤Î¥Ç¡¼¥¿¤ÈÊýË¡¤Ë¾ÇÅÀ¤òÅö¤Æ¤Þ¤·¤¿¡£¥È¥ì¡¼¥Ë¥ó¥°¥Ç¡¼¥¿¤È¤·¤Æ¤Ï²Ê³Ø¤äÆü¾ïÀ¸³è¡¢¿´Íý¤Ê¤É¤Î¾ï¼±¡¦°ìÈÌÃ챤ò¶µ¤¨¤ë¤¿¤á¤Î¥Ç¡¼¥¿¥»¥Ã¥È¤È¡¢¶µ°éŪ²ÁÃͤ䥳¥ó¥Æ¥ó¥Ä¤ÎÉʼÁ¤Ë´ð¤Å¤¤¤Æ¿µ½Å¤Ë¥Õ¥£¥ë¥¿¥ê¥ó¥°¤·¤¿¥¦¥§¥Ö¤Î¥Ç¡¼¥¿¥»¥Ã¥È¤ò»ÈÍѤ·¤¿¤È¤Î¤³¤È¡£¤Þ¤¿¡¢ºÇ½é¤ËPhi-1.5¤ò¥È¥ì¡¼¥Ë¥ó¥°¤·¤Æ¤½¤ÎÃ챤òPhi-2¤Ø°Üž¤¹¤ë¤È¤¤¤¦ÊýË¡¤òÍѤ¤¤ë¤³¤È¤Ç¥È¥ì¡¼¥Ë¥ó¥°¤Î¼ý«¤ò²Ã®¤·¤Ä¤Ä¥Ù¥ó¥Á¥Þ¡¼¥¯¥¹¥³¥¢¤ò¸þ¾å¤µ¤»¤ë¤³¤È¤ËÀ®¸ù¤·¤Þ¤·¤¿¡£¥È¥ì¡¼¥Ë¥ó¥°¥Ç¡¼¥¿¤ÎÎ̤Ï1.4Ãû¥È¡¼¥¯¥ó¤Ç¡¢96¸Ä¤ÎNVIDIA A100 GPU¤ò»ÈÍѤ·¤Æ14Æü¤«¤±¤Æ¥È¥ì¡¼¥Ë¥ó¥°¤ò¹Ô¤Ã¤¿¤È¤Î¤³¤È¡£
²¼¿Þ¤ÏToxiGen¤Ë´ð¤Å¤¤¤Æ°ÂÁ´À¥¹¥³¥¢¤ò»»½Ð¤·¤¿¤â¤Î¤Ç¡¢¥¹¥³¥¢¤¬¹â¤¤¤Û¤É̵³²¤Êʸ¤òÀ¸À®¤·¤ä¤¹¤¤¤³¤È¤ò¼¨¤·¤Æ¤¤¤Þ¤¹¡£Phi-2¤ÏRLHF¤ä¥Õ¥¡¥¤¥ó¥Á¥å¡¼¥Ë¥ó¥°¤Ë¤è¤ëÄ´À°¤ò¹Ô¤Ã¤Æ¤¤¤Ê¤¤¥Ù¡¼¥¹¥â¥Ç¥ë¤Ç¤¢¤ë¤Ë¤â¤«¤«¤ï¤é¤º¡¢ÆÇÀ¤ä¥Ð¥¤¥¢¥¹¤Ë´Ø¤·¤Æ´û¸¤ÎÄ´À°ºÑ¤ß¥â¥Ç¥ë¤Ç¤¢¤ëLlama2-7b¤è¤ê¤âÍ¥¤ì¤¿·ë²Ì¤ò½Ð¤·¤Þ¤·¤¿¡£
¾¤Î¥â¥Ç¥ë¤ÈÈæ³Ó¤·¤¿¥Ù¥ó¥Á¥Þ¡¼¥¯·ë²Ì¤Ï°Ê²¼¤ÎÄ̤ꡣPhi-2¤ÏÆÃ¤Ë¥×¥í¥°¥é¥ß¥ó¥°¤ä¿ô³Ø¤Ê¤ÉÊ£¿ô¤Î¥¹¥Æ¥Ã¥×¤¬É¬ÍפʿäÏÀ¥¿¥¹¥¯¤Ë¤ª¤¤¤ÆÍ¥¤ì¤¿¥Ñ¥Õ¥©¡¼¥Þ¥ó¥¹¤òȯ´ø¤·¤Æ¤¤¤Þ¤¹¡£
¥â¥Ç¥ë¥µ¥¤¥ºBBH¾ï¼±Åª¿äÏÀ¸À¸ìÍý²ò¿ô³Ø¥×¥í¥°¥é¥ß¥ó¥°Llama-27B40.062.256.716.521.013B47.865.061.934.225.470B66.569.267.664.138.3Mistral7B57.266.463.746.439.4Phi-22.7B59.268.862.061.153.7
2023ǯ12·î6Æü¤ËÅо줷¤¿Gemini¤Î¤¦¤Á¡¢°ìÈÖ¾®¤µ¤¤¥â¥Ç¥ë¤Ç¤¢¤ë¡ÖGemini Nano 2¡×¤È¤ÎÈæ³Ó¤Ï²¼¿Þ¤ÎÄ̤ꡣPhi-2¤ÏGemini Nano 2¤ÈƱÅù°Ê¾å¤ÎÀǽ¤ò»ý¤Ã¤Æ¤¤¤ë¤³¤È¤¬³Îǧ¤Ç¤¤Þ¤¹¡£
ModelSizeBBHBoolQMBPPMMLUGemini Nano 23.2B42.479.327.255.8Phi-22.7B59.383.359.156.7
¥Ù¥ó¥Á¥Þ¡¼¥¯Ä̤ꡢñ½ã¤ÊʪÍýÌäÂê¤Ç¤¢¤ì¤ÐÆñ¤Ê¤¯²ò¤±¤ëÌÏÍÍ¡£Ê¿Êýº¬¤Î·×»»¤Ë¤Ä¤¤¤Æ¤â¤Û¤ÜÀµ³Î¤Ë¹Ô¤¦¤³¤È¤¬¤Ç¤¤Þ¤·¤¿¡£
¤Ê¤ª¡¢Phi-2¤Ï¸¦µæÍÑÅӤΤߤΥ饤¥»¥ó¥¹¤ÇÄ󶡤µ¤ì¤Æ¤¤¤ë¤¿¤á¡¢¾¦ÍÑÍøÍѤÏÉÔ²Äǽ¤È¤Ê¤Ã¤Æ¤¤¤ëÅÀ¤ËÃí°Õ¤¬É¬ÍפǤ¹¡£