基于StarCoder2 7B的高效编码Dolphin模型家族的无审查版本。

7B

227 提交 更新于4个月前

说明

DolphinCoder StarCoder2 7b 🐬

赞助商:[latitude.sh](https://www.latitude.sh/).

加入我们的Discord![https://discord.gg/cognitivecomputations](https://discord.gg/cognitivecomputations)

此模型基于StarCoder2-7b,并受bigcode-openrail-m许可协议约束。

这只海豚编程真的很厉害,我使用大量编程数据进行训练。

此模型未经过筛选。我已经过滤了数据集以移除对齐偏差。这使得模型更合规。建议在将模型作为服务发布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。请阅读我关于未筛选模型的博客文章。[https://erichartford.com/uncensored-models](https://erichartford.com/uncensored-models)
您对本模型创建的任何内容负责。请负责任地享受。

训练

使用qLoRA和Axolotl在8x L40S上训练了3个周期,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](https://www.latitude.sh/)的慷慨赞助。
  • 欢迎Microsoft加入开源AI!感谢提供Orca-Math数据集!
  • 对[BigCode](https://www.bigcode-project.org/)表示感谢,它训练并发布了StarCoder2的权重。
  • 衷心感谢数据集的作者:@ise-uiuc, @teknium, @m-a-p
  • 还衷心感谢@winglian以及Axolotl的贡献者,他们打造了最佳训练框架!
  • Built with Axolotl
  • 感谢开源人工智能社区中所有教导我并帮助我前行的人们。

示例输出

如果你希望资助我的努力

周边产品