基于StarCoder2 7B的,擅长编码技术的Dolphin模型家族的不审查版本。

7B

227 拉取 更新于4个月前

说明

DolphinCoder StarCoder2 7b 🐬

latitude.sh 赞助。

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

此模型基于Starcoder2-7b,并根据bigcode-openrail-m许可证受权。

这只海豚在编程方面真的很好,我用很多编程数据训练过它。

此模型未经过审查。我已经过滤了数据集以移除对齐和偏差。这使得模型更符合规范。请在将模型公开为服务之前自行实现对齐层。它将高度遵守任何请求,包括不道德的请求。请阅读我关于未审查模型的博客文章。 https://erichartford.com/uncensored-models
你将对使用此模型创建的任何内容负责。负责任地享受。

训练

使用qLoRA和Axolotl,在8x L40S上训练3个epoch需要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

  • 本模型的实现得益于 latitude.sh 的慷慨赞助。
  • 欢迎微软加入开源AI!感谢您提供的Orca-Math数据集!
  • 衷心感谢 BigCode 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
  • 向@winglian和Axolotl的贡献者表示极大的感谢,他们创造了最好的训练框架!
  • Built with Axolotl
  • 感谢所有在开源人工智能社区中教导我并一路帮助我的人。

示例输出

如果您想资助我的努力

周边产品