{"id":1061,"date":"2024-04-30T00:15:00","date_gmt":"2024-04-29T15:15:00","guid":{"rendered":"https:\/\/staka.jp\/wordpress\/?p=1061"},"modified":"2024-04-30T00:15:00","modified_gmt":"2024-04-29T15:15:00","slug":"llama-3-gemma-recurrent-gemma-phi-3%e7%ad%89%e3%81%ae%e6%a9%9f%e6%a2%b0%e7%bf%bb%e8%a8%b3%e6%80%a7%e8%83%bd","status":"publish","type":"post","link":"https:\/\/staka.jp\/wordpress\/?p=1061","title":{"rendered":"Gemma 7B, Llama 3 8B, Phi-3 mini, recurrent gemma 2B \u7b49\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd"},"content":{"rendered":"\n

\u5c0f\u898f\u6a21\u3067\u3082\u6027\u80fd\u304c\u826f\u3044\u3068\u4e3b\u5f35\u3059\u308b\u30e2\u30c7\u30eb\u304c\u6b21\u3005\u3068\u767a\u8868\u3055\u308c\u3066\u3044\u308b\uff08OSS \u2013 arXiv\u6700\u65b0\u8ad6\u6587\u306e\u7d39\u4ecb (devneko.jp)<\/a>\uff09\u30027B\u304b\u30898B\u306e\u30e2\u30c7\u30eb\u306fGPU 1\u3064\u3067\u3082\u5b9f\u884c\u3067\u304d\u4f7f\u3044\u52dd\u624b\u304c\u826f\u3044\u3002\u4e00\u65b9\u3067\u904e\u53bb\u306e\u691c\u8a3c\u3067\u306f\u5341\u5206\u306a\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd\u3092\u51fa\u3059\u3053\u3068\u304c\u96e3\u3057\u305d\u3046\u3067\u3042\u308b [1]<\/p>\n\n\n\n

\u6700\u65b0\u30e2\u30c7\u30eb\u3067\u4e0a\u8a18\u5236\u7d04\u304c\u89e3\u6d88\u3057\u3066\u3044\u308b\u306e\u304b\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd\u3092\u691c\u8a3c\u3057\u3066\u307f\u305f\u3002\u6027\u80fd\u8a55\u4fa1\u306b\u4f7f\u7528\u3057\u305f\u30c7\u30fc\u30bf\u306f\u4ee5\u524d\uff08DAMO PolyLM-13B\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd | \u3077\u308b\u30fc\u3075\u304a\u3076\u3053\u3093\u305b\u3077\u3068 (staka.jp)<\/a>\uff09\u3068\u540c\u3058\u3002\u691c\u8a3c\u74b0\u5883\u306f\u74b0\u5883\u306fColab Pro+\u3092\u7528\u3044\u305f\u3002\u57fa\u672c\u7684\u306bfp16\u3067\u6271\u3063\u3066\u3044\u308b\u3002<\/p>\n\n\n\n

\u672c\u4ef6\u306e\u691c\u8a3c\u3067\u306f\u624b\u3067\u56de\u7b54\u62bd\u51fa\u3092\u3059\u308b\u4ee3\u308f\u308a\u306bGPT-4 Turbo\u3092\u7528\u3044\u3066\u56de\u7b54\u62bd\u51fa\u3092\u884c\u3063\u305f\u7d50\u679c\u3092\u4f75\u8a18\u3057\u305f\u3002\u8a55\u4fa1\u6307\u6a19\u306fBLEU\u3001\u4f7f\u7528\u3057\u305f\u30c4\u30fc\u30eb\u3084tokenizer\u306f\u4ee5\u524d\u3068\u540c\u3058\uff08sacrebleu \u2013tokenize ja-mecab\uff09\u3067\u3042\u308b\u3002<\/p>\n\n\n\n

\u300c0 shot\u300d\u3068\u300c1 shot\u300d\u306e\u6bd4\u8f03\u3067ICL\u3084RAG\u306a\u3069\u30d7\u30ed\u30f3\u30d7\u30c8\u5185\u306b\u30c7\u30fc\u30bf\u3092\u4e0e\u3048\u305f\u6642\u306e\u6027\u80fd\u3001\u300c\u51fa\u529b\u305d\u306e\u307e\u307e\u300d\u3068\u300cGPT-4\u3067\u56de\u7b54\u62bd\u51fa\u300d\u306e\u6bd4\u8f03\u3067\u6307\u793a\u306e\u3057\u3084\u3059\u3055\u3092\u30b6\u30c3\u30af\u30ea\u3068\u307f\u308b\u4e8b\u304c\u3067\u304d\u308b[2]\u3002<\/p>\n\n\n\n

\u30e2\u30c7\u30eb\u30fb\u6761\u4ef6<\/td>param<\/td>0 shot \/
\u51fa\u529b\u305d\u306e\u307e\u307e<\/td>		0 shot \/
GPT-4\u3067\u56de\u7b54\u62bd\u51fa<\/td>		1 shot \/
\u51fa\u529b\u305d\u306e\u307e\u307e<\/td>		1 shot \/
GPT-4\u3067\u56de\u7b54\u62bd\u51fa<\/td><\/tr>
google\/gemma-1.1-7b-it \u00b7 Hugging Face<\/a><\/td>8.54B<\/td>17.1<\/span><\/td>17.1<\/span><\/td>32.5<\/strong><\/td>32.2<\/strong><\/td><\/tr>
meta-llama\/Meta-Llama-3-8B-Instruct \u00b7 Hugging Face<\/a><\/td>8.03B<\/td>11.9<\/td>16.2<\/td>20.6<\/td>30.2<\/span><\/td><\/tr>
Qwen\/Qwen1.5-7B-Chat \u00b7 Hugging Face<\/a><\/td>7.72B<\/td>17.8<\/strong><\/td>17.9<\/strong><\/td>18.9<\/td>19.0<\/td><\/tr>
microsoft\/Phi-3-mini-128k-instruct \u00b7 Hugging Face<\/a><\/td>3.82B<\/td>12.6<\/td>12.7<\/td>20.0<\/td>20.3<\/td><\/tr>
google\/recurrentgemma-2b-it \u00b7 Hugging Face<\/a><\/td>2.68B<\/td>7.9<\/td>8.7<\/td>20.3<\/td>20.9<\/td><\/tr>
\uff08\u53c2\u8003\uff09gpt-3.5-turbo<\/td><\/td>27.1<\/td>26.9<\/td>37.1<\/td>37.0<\/td><\/tr><\/tbody><\/table>
\u5404\u7a2e\u5c0f\u898f\u6a21\u30e2\u30c7\u30eb\u306eBLEU<\/figcaption><\/figure>\n\n\n\n

\u3088\u308a\u5927\u898f\u6a21\u306a\u30e2\u30c7\u30eb\u306b\u3064\u3044\u3066\u3082\u691c\u8a3c\u3092\u884c\u3063\u305f\u3002\u3053\u3061\u3089\u306follama\u3092\u7528\u3044\u3066\u691c\u8a3c\u3092\u884c\u3063\u3066\u3044\u308b\u3002<\/p>\n\n\n\n

\u5b9f\u306f\u4e0a\u8a18\u8868\u306e\u30e2\u30c7\u30eb\u3082\u5f53\u521dollama\u3092\u7528\u3044\u3066\u691c\u8a3c\u3092\u884c\u3063\u3066\u3044\u305f\u304c\u6027\u80fd\u304c\u60aa\u304b\u3063\u305f\u30024bit\u91cf\u5b50\u5316\u306e\u5f71\u97ff\u3067\u306f\u306a\u3044\u304b\u3068\u8003\u3048\u3066\u3044\u308b[3]\u304c\u8a73\u7d30\u306f\u4e0d\u660e\u3067\u3042\u308b\u3002\u4e0b\u8868\u306e\u7d50\u679c\u3082fp16\u3067\u5b9f\u884c\u3057\u305f\u5834\u5408\u3001\u3088\u308a\u826f\u3044\u7d50\u679c\u306b\u306a\u308b\u53ef\u80fd\u6027\u304c\u3042\u308b\u3002<\/p>\n\n\n\n

\u30e2\u30c7\u30eb\u30fb\u6761\u4ef6<\/td>param<\/td>0 shot \/
\u51fa\u529b\u305d\u306e\u307e\u307e<\/td>		0 shot \/
GPT-4\u3067\u56de\u7b54\u62bd\u51fa<\/td>		1 shot \/
\u51fa\u529b\u305d\u306e\u307e\u307e<\/td>		1 shot \/
GPT-4\u3067\u56de\u7b54\u62bd\u51fa<\/td><\/tr>
command-r-plus (ollama.com)<\/a><\/td>104B<\/td>27.4<\/strong><\/td>29.4<\/strong><\/td>37.1<\/strong><\/td>36.8<\/span><\/td><\/tr>
llama3:70b-instruct (ollama.com)<\/a><\/td>71B<\/td>17.3<\/td>16.5<\/td>21.0<\/td>32.3<\/td><\/tr>
\uff08\u53c2\u8003\uff09gpt-3.5-turbo<\/td><\/td>27.1<\/td>26.9<\/td>37.1<\/strong><\/td>37.0<\/strong><\/td><\/tr><\/tbody><\/table>
\u5404\u7a2e\u5927\u898f\u6a21\u30e2\u30c7\u30eb\u306eBLEU<\/figcaption><\/figure>\n\n\n\n

\u7d50\u679c\u3068\u7d50\u8ad6<\/h2>\n\n\n\n

Llama 3 8B, Gemma 7B\u306f\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u306e\u5272\u306b\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd\u304c\u9ad8\u3044\u3002\u7279\u306bGemma\u306f\u65e5\u672c\u8a9e\u5bfe\u5fdc\u3001\u547d\u4ee4\u306e\u7406\u89e3\u5171\u306b\u6027\u80fd\u304c\u9ad8\u3044\u3002Llama 3 8B Instruct\u306f\u51fa\u529b\u306b\u4e0d\u8981\u306a\u90e8\u5206\uff08\u4eca\u304b\u3089\u7ffb\u8a33\u3057\u307e\u3059\u306a\u3069\u306e\u30ea\u30fc\u30c9\u6587\uff09\u304c\u51fa\u308b\u4e8b\u304c\u591a\u304f\u305d\u308c\u304c\u898b\u305f\u76ee\u306eBLEU\u306b\u5f71\u97ff\u3057\u3066\u3044\u305f\u3002\u4e3b\u305f\u308b\u6587\u7ae0\u306e\u307f\u3092\u629c\u304d\u51fa\u305b\u3070\u6027\u80fd\u306f\u60aa\u304f\u306a\u3044\u3002\u65e5\u672c\u8a9e\u3092\u7d99\u7d9a\u5b66\u7fd2\u3057\u305f\u30d0\u30fc\u30b8\u30e7\u30f3\u3084\u65e5\u672c\u8a9e\u306e\u547d\u4ee4\u30bb\u30c3\u30c8\u3067SFT\u3057\u305f\u30d0\u30fc\u30b8\u30e7\u30f3\u306b\u671f\u5f85\u3067\u3042\u308b\u3002<\/p>\n\n\n\n

\u3088\u308a\u5c0f\u578b\u306ePhi-3\u3001reccurent gemma 2B\u306f\u5b9f\u7528\u306b\u306f\u53b3\u3057\u305d\u3046\u3060\u304c\u671f\u5f85\u306e\u3067\u304d\u308b\u7d50\u679c\u3092\u51fa\u3057\u3066\u3044\u308b\u3088\u3046\u306b\u601d\u3046\u3002\u7279\u306breccurent gemma\u306e\u3088\u3046\u306b\u65b0\u305f\u306a\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3[4]\u306b\u306f\u671f\u5f85\u5927\u3067\u3042\u308b\u3002\u6a5f\u68b0\u7ffb\u8a33\u3067\u306fSSM\u306e\u3088\u3046\u306a\u72b6\u614b\u3092\u6271\u3048\u308b\u30e2\u30c7\u30eb\u3060\u3068\u5229\u70b9\u304c\u591a\u3044\u3002\u4f8b\u3048\u3070\u300c\u5358\u8a9e\u5bfe\u5fdc\u306e\u8f9e\u66f8\u300d\u3084\u300c\u8a33\u8abf\u300d\u3001\u300c\u5b9a\u578b\u6587\u300d\u306a\u3069\u3092\u72b6\u614b\u306b\u62bc\u3057\u8fbc\u3093\u3067\u304a\u304f\u3053\u3068\u3067\u3001\u6b32\u3057\u3044\u30b9\u30bf\u30a4\u30eb\u306e\u6a5f\u68b0\u7ffb\u8a33\u3092\u5b9f\u884c\u53ef\u80fd\u306b\u306a\u308b\u3002FuguMT ver 2\u306f\u3053\u306e\u3088\u3046\u306a\u6a5f\u80fd\u3067\u69cb\u7bc9\u3059\u308b\u4e88\u5b9a\u3067\u3042\u308b\u3002<\/p>\n\n\n\n

\u3088\u308a\u5927\u898f\u6a21\u306a\u30e2\u30c7\u30eb\u3068\u3057\u3066\u306fCohere\u306eCommand R+\u306f\u3055\u3059\u304c\u306e\u6027\u80fd\u3067\u3042\u308b[5]\u3002Llama 3 70B\u306b\u3064\u3044\u3066\u306f\u91cf\u5b50\u5316\u306e\u5f71\u97ff\u3092\u53d7\u3051\u3066\u3044\u305d\u3046\u306a\u306e\u3067fp16\u3067\u306e\u6027\u80fd\u3092\u8a66\u3057\u3066\u307f\u305f\u3044\u3068\u3053\u308d\u3002<\/p>\n\n\n\n

FuguMT ver2\uff08\u8f9e\u66f8\u30fb\u7ffb\u8a33\u30b9\u30bf\u30a4\u30eb\u6307\u5b9a\u53ef\u80fd\u306aLLM\u30d9\u30fc\u30b9\u306e\u6a5f\u68b0\u7ffb\u8a33\u30e2\u30c7\u30eb\uff09\u516c\u958b\u306b\u5408\u308f\u305b\u3066\u3001\u3053\u308c\u3089\u691c\u8a3c\u30c7\u30fc\u30bf\u306e\u6700\u65b0\u5316\u3084\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u306e\u516c\u958b\u3082\u3057\u305f\u3044[6]\u3002<\/p>\n\n\n\n

\u811a\u6ce8<\/h4>\n\n\n\n

[1] \u4ee5\u524d\u306e\u691c\u8a3c\u3067\u306f7B\u30af\u30e9\u30b9\u306814B\u30af\u30e9\u30b9\u3067\u5927\u304d\u306a\u5dee\u304c\u51fa\u3066\u3044\u305f\u3002PaLM2 \/ ELYZA-japanese-Llama-2-7b\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd | \u3077\u308b\u30fc\u3075\u304a\u3076\u3053\u3093\u305b\u3077\u3068 (staka.jp)<\/a>
[2] \u30e2\u30c7\u30eb\u306b\u5408\u308f\u305b\u305f\u30d7\u30ed\u30f3\u30d7\u30c8\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u306f\u884c\u3063\u3066\u3044\u306a\u3044\u305f\u3081\u3001\u7279\u306b\u5f8c\u8005\u306f\u672c\u5f53\u306b\u30b6\u30c3\u30af\u30ea\u3068\u3067\u3042\u308b\u3002
[3] Llama3\u306e\u91cf\u5b50\u5316\u306e\u5f71\u97ff\u306b\u3064\u3044\u3066\u306e\u691c\u8a3c\u306f GitHub – Macaronlin\/LLaMA3-Quantization<\/a> \u304c\u8a73\u3057\u3044\u3002
[4] Fugu-MT \u8ad6\u6587\u7ffb\u8a33(\u6982\u8981): RecurrentGemma: Moving Past Transformers for Efficient Open Language Models (fugumt.com)<\/a>\u3001Griffin\u30a2\u30fc\u30ad\u30c6\u30af\u30c1\u30e3Fugu-MT \u8ad6\u6587\u7ffb\u8a33(\u6982\u8981): Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models (fugumt.com)<\/a>
[5] Command R+ (cohere.com)<\/a>\u3001Introducing Command R+: A Scalable LLM Built for Business (cohere.com)<\/a>\u3001\u7814\u7a76\u76ee\u7684\u7528\u306bCC-BY-NC-4.0\u3067\u516c\u958b\u3055\u308c\u3066\u3044\u308b\u5546\u7528\u30e2\u30c7\u30eb\u3067\u3042\u308b\u306e\u3067\u5f53\u305f\u308a\u524d\u304b\u3082\u3057\u308c\u306a\u3044\u304c\u3002\u3002
[6] \u3060\u304c\u3001\u6642\u9593\u306e\u4f59\u88d5\u304c\u306a\u3044\u72b6\u6cc1\u304c\u7d9a\u304f\u3002\u65e2\u5b58FuguMT\u306e\u554f\u984c\u4fee\u6b63\u3082\u3044\u3044\u52a0\u6e1b\u3084\u308a\u305f\u3044\u3002\u3002\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u5c0f\u898f\u6a21\u3067\u3082\u6027\u80fd\u304c\u826f\u3044\u3068\u4e3b\u5f35\u3059\u308b\u30e2\u30c7\u30eb\u304c\u6b21\u3005\u3068\u767a\u8868\u3055\u308c\u3066\u3044\u308b\uff08OSS \u2013 arXiv\u6700\u65b0\u8ad6\u6587\u306e\u7d39\u4ecb (devneko.jp)\uff09\u30027B\u304b\u30898B\u306e\u30e2\u30c7\u30eb\u306fGPU 1\u3064\u3067\u3082\u5b9f\u884c\u3067\u304d\u4f7f\u3044\u52dd\u624b\u304c\u826f\u3044\u3002\u4e00\u65b9\u3067\u904e\u53bb\u306e\u691c\u8a3c\u3067\u306f\u5341\u5206\u306a\u6a5f\u68b0\u7ffb … “Gemma 7B, Llama 3 8B, Phi-3 mini, recurrent gemma 2B \u7b49\u306e\u6a5f\u68b0\u7ffb\u8a33\u6027\u80fd” \u306e<\/span>\u7d9a\u304d\u3092\u8aad\u3080<\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2,13],"tags":[35,37,55],"class_list":["post-1061","post","type-post","status-publish","format-standard","hentry","category-ai","category-13","tag-gemma","tag-gpt-3-5-turbo","tag-phi-3"],"_links":{"self":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/1061","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1061"}],"version-history":[{"count":0,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=\/wp\/v2\/posts\/1061\/revisions"}],"wp:attachment":[{"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1061"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1061"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/staka.jp\/wordpress\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1061"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}