基于 StarCoder2 7B 的编码能力极强的未美化版本 Dolphin 模型家族。

7B

227 提取 更新 4 个月前

说明読み書き

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 许可证约束。

Dolphin 的编程能力非常出色,我使用了很多编程数据来训练。

此模型未被审查。我已过滤数据集以去除对齐和偏见。这使模型更合规。建议在使用此模型作为服务之前实现自己的对齐层。它将对任何请求高度合规,甚至对不道德的请求。请阅读我关于未审查模型的博客文章。 https://erichartford.com/uncensored-models
您对使用此模型创建的任何内容负责任。负责任地享受。

训练

使用 qLoRA 和 Axolotl 在 8x L40S 上训练 3 个周期花费了 2 天时间。

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

  • 这个模型得益于 latitude.sh 的慷慨赞助。
  • 欢迎微软加入开源人工智能!感谢 Orca-Math 数据集!
  • 感谢<协会>BigCode提供训练和发布StarCoder2的权重。
  • 向数据集的作者致以最诚挚的谢意:@ise-uiuc、@teknium、@m-a-p。
  • 并对@winglian和Axolotl的贡献者表示由衷的感谢,他们打造了最优秀的训练框架!
  • Built with Axolotl
  • 感谢开源AI社区中所有其他人在我学习和成长过程中给予的帮助和支持。

示例输出

如果您想财务支持我的工作

周边产品