3.专业8B语言模型：这是一个针对科学文献综合任务优化的8B参数语言模型，在性能与计算效率之间取得了很好平衡。团队基于来自迭代自我反馈生成管道生成的合成数据训练并微调了Llama 3.1 8B。

对比基于 OpenAI 的 GPT-4o 和 Anthropic 的 Claude 的科研模型可以发现，它们虽然性能很强，但价格昂贵、而且工作原理也不透明。而 OpenScholar 正是为了挑战这些现有的大模型玩家入局的！

新智元报道  编辑：alan【新智元导读】代码模型可以自己进化，利用自身生成的数据来进行指令调优，效果超越GPT-4o直接蒸馏！LLM作为智能的基座，可以衍生出各种能力。代码能力就是其中一种：程序补全、注释、优化、修bug、测试等等。而想要充分发挥LLM的巨大潜力，指令调优（Instruction ...

and games from scratch — all of which are made more powerful with GPT-4, of course. A recent benchmark test found that both of OpenAI’s newest models, o1-preview and o1-mini, can code with ...

Also: 6 ways OpenAI just supercharged ChatGPT for free users The model is 50% cheaper in OpenAI's API than GPT-4 Turbo while still matching its English and coding capabilities and outperforming ...

In nine out of 12 such evaluations, Qwen2.5 Coder’s flagship variant performed better than GPT-4o and Claude 3.5 Sonnet, according to the statement. Until now, the coding capabilities of open ...

Alvaro Cintra used GPT-4o to generate Python code for a fully working video game called ‘Breakout,’ starting from just a screenshot of the game and the simple prompt, "Can you please code this ...

In nine out of 12 evaluations, Qwen2.5 Coder’s flagship variant performed better than GPT-4o and Claude 3.5 Sonnet, according ...

With improved reasoning and coding capabilities, they establish a standard that challenges even leading models like GPT-4 and Google Gemini. What truly distinguishes them, however, is their ...

奇月 发自 凹非寺量子位 | 公众号 QbitAI 只需几秒钟，开源模型检索4500篇论文，比GPT-4o还靠谱！ 这就是由华盛顿大学和艾伦人工智能研究所（Ai2）打造的最新模型OpenScholar。 它还是首个从论文到数据集、模型检查点都完全开源的科研助手模型。 在由20位专家进行的500次对比实验中，72%的情况下他们都觉得OpenScholar的输出结果超越了人类。 而且OpenSchol ...