基于StarCoder2 7B的编码表现卓越的Dolphin模型系列的无审查版本。

7B

227 Pulls 更新于4个月前

自述文件

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许可约束。

这只海豚真的擅长编程,我用很多编程数据进行了训练。

此模型未经审查。我对数据集进行了过滤,以去除对齐和偏见。这使模型更加合规。请在将模型作为服务公开之前实现自己的对齐层。它将对任何请求高度合规,甚至包括不道德的请求。请阅读我的关于未经审查的模型的博客文章。https://erichartford.com/uncensored-models
您对此模型创建的任何内容负责。负责任地享受。

训练

在8x L40S的机器上,使用qLoRA和Axolotl训练了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

感激

  • 本模型得以实现,得益于latitude.sh的慷慨赞助。
  • 欢迎微软加入开源AI!感谢提供Orca-Math数据集!
  • BigCode表示感谢,它训练并发布了StarCoder2的权重。
  • 对数据集作者@ise-uiuc, @teknium, @m-a-p表示巨大感谢。
  • 对@winglian和Axolotl的贡献者表示巨大感谢,他们制作了最好的训练框架!
  • Built with Axolotl
  • 感谢开源AI社区中所有在途中教导我的和帮助过我的人。

示例输出

如果您想资助我的工作进行

周边产品