{"id":18480,"date":"2025-01-15T19:08:48","date_gmt":"2025-01-15T11:08:48","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=18480"},"modified":"2025-01-15T19:23:16","modified_gmt":"2025-01-15T11:23:16","slug":"codestral","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/codestral\/","title":{"rendered":"Mistral AI \u53d1\u5e03 Codestral 25.01 \u6a21\u578b\uff1a\u66f4\u5927\u3001\u66f4\u5feb\u3001\u66f4\u5f3a\u7684\u7f16\u7a0b\u4e13\u7528\u6a21\u578b"},"content":{"rendered":"

\u4ee5 Tab \u952e\u7684\u901f\u5ea6\u7f16\u5199\u4ee3\u7801\u3002\u4eca\u5929\u5728 Continue.dev<\/a> \u4e2d\u53ef\u7528\uff0c\u5f88\u5feb\u5c06\u5728\u5176\u4ed6\u9886\u5148\u7684 AI \u4ee3\u7801\u52a9\u624b\u4e0a\u63a8\u51fa\u3002<\/p>\n

\u5728\u8fc7\u53bb\u4e00\u5e74\u4e2d\uff0cAI \u7684\u6240\u6709\u521b\u65b0\u4e2d\uff0c\u4ee3\u7801\u751f\u6210\u53ef\u4ee5\u8bf4\u662f\u6700\u91cd\u8981\u7684\u3002\u7c7b\u4f3c\u4e8e\u6d41\u6c34\u7ebf\u7b80\u5316\u5236\u9020\u4e1a\u548c\u8ba1\u7b97\u5668\u6539\u53d8\u6570\u5b66\u7684\u65b9\u5f0f\uff0c\u7f16\u7801\u6a21\u578b\u4ee3\u8868\u4e86\u8f6f\u4ef6\u5f00\u53d1\u7684\u91cd\u5927\u53d8\u9769\u3002<\/p>\n

Mistral AI<\/a> \u4e00\u76f4\u5904\u4e8e\u8fd9\u4e00\u53d8\u9769\u7684\u524d\u6cbf\uff0c\u63a8\u51fa\u4e86\u00a0Codestral\uff0c\u8fd9\u662f\u4e00\u4e2a\u4eca\u5e74\u65e9\u4e9b\u65f6\u5019\u53d1\u5e03\u7684\u5148\u8fdb\uff08SOTA\uff09\u7f16\u7801\u6a21\u578b\u3002Codestral \u8f7b\u91cf\u7ea7\u3001\u5feb\u901f\uff0c\u7cbe\u901a 80 \u591a\u79cd\u7f16\u7a0b\u8bed\u8a00\uff0c\u9488\u5bf9\u4f4e\u5ef6\u8fdf\u3001\u9ad8\u9891\u7387\u7528\u4f8b\u8fdb\u884c\u4e86\u4f18\u5316\uff0c\u5e76\u652f\u6301\u4e2d\u95f4\u586b\u5145\uff08FIM\uff09\u3001\u4ee3\u7801\u66f4\u6b63\u548c\u6d4b\u8bd5\u751f\u6210\u7b49\u4efb\u52a1\u3002Codestral \u5df2\u88ab\u6210\u5343\u4e0a\u4e07\u7684\u5f00\u53d1\u4eba\u5458\u7528\u4f5c\u529f\u80fd\u5f3a\u5927\u7684\u7f16\u7801\u52a9\u624b\uff0c\u7ecf\u5e38\u4f7f\u751f\u4ea7\u529b\u63d0\u9ad8\u6570\u500d\u3002\u4eca\u5929\uff0cCodestral \u6b63\u5728\u8fdb\u884c\u91cd\u5927\u5347\u7ea7\u3002<\/p>\n

Codestral 25.01<\/strong>\u00a0\u91c7\u7528\u4e86\u6bd4\u539f\u59cb\u7248\u672c\u66f4\u9ad8\u6548\u7684\u67b6\u6784\u548c\u6539\u8fdb\u7684 Tokenizer\uff0c\u751f\u6210\u548c\u5b8c\u6210\u4ee3\u7801\u7684\u901f\u5ea6\u5feb\u7ea6 2 \u500d\u3002\u8be5\u6a21\u578b\u73b0\u5728\u662f\u5176\u6743\u91cd\u7ea7\u522b\u4e2d\u4ee3\u7801\u7f16\u5199\u7684\u660e\u663e\u9886\u5bfc\u8005\uff0c\u5e76\u4e14\u5728 FIM \u7528\u4f8b\u4e2d\u5168\u9762\u5904\u4e8e SOTA \u6c34\u5e73\u3002<\/p>\n

\u57fa\u51c6\u6d4b\u8bd5<\/strong><\/h3>\n

\u6211\u4eec\u4f7f\u7528\u88ab\u5e7f\u6cdb\u8ba4\u4e3a\u662f FIM \u4efb\u52a1\u6700\u4f73\u7684\u9886\u5148\u7684\u53c2\u6570\u5c0f\u4e8e 100B \u7684\u7f16\u7801\u6a21\u578b\u5bf9\u65b0\u7684 Codestral \u8fdb\u884c\u4e86\u57fa\u51c6\u6d4b\u8bd5\u3002<\/p>\n

\u6982\u8ff0<\/strong><\/p>\n\n\n\n\n\n\n\n\n\n\n
<\/th>\n<\/th>\nPython<\/th>\nSQL<\/th>\n<\/th>\n\u591a\u79cd\u8bed\u8a00\u7684\u5e73\u5747\u503c<\/th>\n<\/tr>\n<\/thead>\n
\u6a21\u578b<\/strong><\/td>\n\u4e0a\u4e0b\u6587\u957f\u5ea6<\/strong><\/td>\nHumanEval<\/strong><\/td>\nMBPP<\/strong><\/td>\nCruxEval<\/strong><\/td>\nLiveCodeBench<\/strong><\/td>\n<\/tr>\n
Codestral-2501<\/td>\n256k<\/strong><\/td>\n86.6%<\/strong><\/td>\n80.2%<\/td>\n55.5%<\/strong><\/td>\n37.9%<\/strong><\/td>\n<\/tr>\n
Codestral-2405 22B<\/td>\n32k<\/td>\n81.1%<\/td>\n78.2%<\/td>\n51.3%<\/td>\n31.5%<\/td>\n<\/tr>\n
Codellama 70B instruct<\/td>\n4k<\/td>\n67.1%<\/td>\n70.8%<\/td>\n47.3%<\/td>\n20.0%<\/td>\n<\/tr>\n
DeepSeek<\/a> Coder 33B instruct<\/td>\n16k<\/td>\n77.4%<\/td>\n80.2%<\/td>\n49.5%<\/td>\n27.0%<\/td>\n<\/tr>\n
DeepSeek Coder V2 lite<\/td>\n128k<\/td>\n83.5%<\/td>\n83.2%<\/strong><\/td>\n49.7%<\/td>\n28.1%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u6309\u8bed\u8a00\u5212\u5206<\/strong><\/p>\n\n\n\n\n\n\n\n\n\n
\u6a21\u578b<\/th>\nHumanEval Python<\/th>\nHumanEval C++<\/th>\nHumanEval Java<\/th>\nHumanEval Javascript<\/th>\nHumanEval Bash<\/th>\nHumanEval Typescript<\/th>\nHumanEval C#<\/th>\nHumanEval\uff08\u5e73\u5747\u503c\uff09<\/th>\n<\/tr>\n<\/thead>\n
Codestral-2501<\/td>\n86.6%<\/strong><\/td>\n78.9%<\/strong><\/td>\n72.8%<\/td>\n82.6%<\/strong><\/td>\n43.0%<\/strong><\/td>\n82.4%<\/strong><\/td>\n53.2%<\/strong><\/td>\n71.4%<\/strong><\/td>\n<\/tr>\n
Codestral-2405 22B<\/td>\n81.1%<\/td>\n68.9%<\/td>\n78.5%<\/strong><\/td>\n71.4%<\/td>\n40.5%<\/td>\n74.8%<\/td>\n43.7%<\/td>\n65.6%<\/td>\n<\/tr>\n
Codellama 70B instruct<\/td>\n67.1%<\/td>\n56.5%<\/td>\n60.8%<\/td>\n62.7%<\/td>\n32.3%<\/td>\n61.0%<\/td>\n46.8%<\/td>\n55.3%<\/td>\n<\/tr>\n
DeepSeek Coder 33B instruct<\/td>\n77.4%<\/td>\n65.8%<\/td>\n73.4%<\/td>\n73.3%<\/td>\n39.2%<\/td>\n77.4%<\/td>\n49.4%<\/td>\n65.1%<\/td>\n<\/tr>\n
DeepSeek Coder V2 lite<\/td>\n83.5%<\/td>\n68.3%<\/td>\n65.2%<\/td>\n80.8%<\/td>\n34.2%<\/td>\n82.4%<\/strong><\/td>\n46.8%<\/td>\n65.9%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

FIM\uff08\u5355\u884c\u7cbe\u786e\u5339\u914d\uff09<\/strong><\/p>\n\n\n\n\n\n\n\n\n\n\n
\u6a21\u578b<\/th>\nHumanEvalFIM Python<\/th>\nHumanEvalFIM Java<\/th>\nHumanEvalFIM JS<\/th>\nHumanEvalFIM\uff08\u5e73\u5747\u503c\uff09<\/th>\n<\/tr>\n<\/thead>\n
Codestral-2501<\/td>\n80.2%<\/strong><\/td>\n89.6%<\/strong><\/td>\n87.96%<\/strong><\/td>\n85.89%<\/strong><\/td>\n<\/tr>\n
Codestral-2405 22B<\/td>\n77.0%<\/td>\n83.2%<\/td>\n86.08%<\/td>\n82.07%<\/td>\n<\/tr>\n
OpenAI FIM API*<\/td>\n80.0%<\/td>\n84.8%<\/td>\n86.5%<\/td>\n83.7%<\/td>\n<\/tr>\n
DeepSeek Chat API<\/td>\n78.8%<\/td>\n89.2%<\/td>\n85.78%<\/td>\n84.63%<\/td>\n<\/tr>\n
DeepSeek Coder V2 lite<\/td>\n78.7%<\/td>\n87.8%<\/td>\n85.90%<\/td>\n84.13%<\/td>\n<\/tr>\n
DeepSeek Coder 33B instruct<\/td>\n80.1%<\/td>\n89.0%<\/td>\n86.80%<\/td>\n85.3%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

FIM pass@1:<\/strong><\/p>\n\n\n\n\n\n\n\n\n
\u6a21\u578b<\/th>\nHumanEvalFIM Python<\/th>\nHumanEvalFIM Java<\/th>\nHumanEvalFIM JS<\/th>\nHumanEvalFIM\uff08\u5e73\u5747\u503c\uff09<\/th>\n<\/tr>\n<\/thead>\n
Codestral-2501<\/td>\n92.5%<\/strong><\/td>\n97.1%<\/strong><\/td>\n96.1%<\/strong><\/td>\n95.3%<\/strong><\/td>\n<\/tr>\n
Codestral-2405 22B<\/td>\n90.2%<\/td>\n90.1%<\/td>\n95.0%<\/td>\n91.8%<\/td>\n<\/tr>\n
OpenAI FIM API*<\/td>\n91.1%<\/td>\n91.8%<\/td>\n95.2%<\/td>\n92.7%<\/td>\n<\/tr>\n
DeepSeek Chat API<\/td>\n91.7%<\/td>\n96.1%<\/td>\n95.3%<\/td>\n94.4%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n