基于StarCoder2 7B,擅长编码的无审查版本的Dolphin模型系列。

7B

227多次提取 更新于4个月前

说明文件

DolphinCoder StarCoder2 7b 🐬

latitude.sh赞助。

加入我们的Discord!https://discord.gg/cognitivecomputations

此模型基于Starcoder2-7b,并受bigcode-openrail-m许可证的约束。

这头海豚在编程方面真的很厉害,我用大量的编程数据进行了训练。

此模型未经审查。我已经过滤了数据集,以删除对齐和偏见。这使得模型更加合规。建议在将模型作为服务公开之前先实现您自己的对齐层。它将高度遵守任何请求,甚至是非道德的请求。请阅读我关于未经审查的模型的博客文章。https://erichartford.com/uncensored-models
您负责使用此模型创建的任何内容。请负责任地享受。

训练

使用qLoRA和Axolotl在8x L40S上训练3个epoch耗时2天

提示格式
此模型使用ChatML提示格式。

<|im_start|>system
You are DolphinCoder, a helpful AI programming assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

示例

<|im_start|>system
You are DolphinCoder, a master at software engineering and coding in any programming language.
<|im_start|>user
Please write me a program in golang that parses all the lines in a file, and reverses them character-wise, and saves it to a new file.
<|im_start|>assistant

量化模型

感谢

  • 此模型得以实现,得益于latitude.sh的慷慨赞助。
  • 欢迎微软进入开源AI!感谢提供Orca-Math数据集!
  • BigCode表示感谢,为训练和发布Starcoder2的权重。
  • 向数据集作者表示衷心的感谢:@ise-uiuc, @teknium, @m-a-p
  • 并衷心感谢@winglian和Axolotl的贡献者,为大家打造了最佳的训练框架!
  • Built with Axolotl
  • 感谢开源人工智能社区中的所有人,感谢他们的教导以及在成长过程中的帮助。

示例输出

如果您想资助我的工作

有偿内容