Recent Releases of chinese-mixtral

chinese-mixtral - 中文Mixtral大模型 v1.2

本次更新添加了仿OpenAI API Demo。教程:https://github.com/ymcui/Chinese-Mixtral/wiki/openaiapizh

This release adds OpenAI API Demo. Tutorial: https://github.com/ymcui/Chinese-Mixtral/wiki/openaiapien

What's Changed

  • Add OpenAI API Demo by @ymcui in https://github.com/ymcui/Chinese-Mixtral/pull/25

Full Changelog: https://github.com/ymcui/Chinese-Mixtral/compare/v1.1...v1.2

- Python
Published by ymcui almost 2 years ago

chinese-mixtral - 中文Mixtral大模型 v1.1

本次更新主要有以下两点:

  • 添加中文Mixtral技术报告,介绍了模型训练方法和相关实验分析

    • 论文地址:https://arxiv.org/abs/2403.01851
  • 添加了预训练和指令精调训练脚本

    • 预训练:https://github.com/ymcui/Chinese-Mixtral/wiki/ptscriptszh
    • 指令精调:https://github.com/ymcui/Chinese-Mixtral/wiki/sftscriptszh

What's Changed

  • Add eval scripts by @iMountTai in https://github.com/ymcui/Chinese-Mixtral/pull/5
  • Update readme and add requirements by @iMountTai in https://github.com/ymcui/Chinese-Mixtral/pull/7
  • llama.cpp: add IQ3_XXS quantization models by @ymcui in https://github.com/ymcui/Chinese-Mixtral/pull/8
  • Add training scripts by @iMountTai in https://github.com/ymcui/Chinese-Mixtral/pull/18
  • Add Chinese Mixtral paper by @ymcui in https://github.com/ymcui/Chinese-Mixtral/pull/20

Full Changelog: https://github.com/ymcui/Chinese-Mixtral/compare/v1.0...v1.1

- Python
Published by ymcui almost 2 years ago

chinese-mixtral - 中文Mixtral大模型 v1.0

发布中文Mixtral, Mixtral-Instruct大模型已正式发布。 - Chinese-Mixtral:基座模型,使用20G语料增量训练 - Chinese-Mixtral-Instruct:指令/chat模型,在Chinese-Mixtral的基础上进一步通过指令精调(500万条指令)获得

模型特点

📖 稀疏混合专家模型

Mixtral是一个稀疏混合专家模型。该模型与以往的LLaMA等主流大模型结构具有显著差异,主要体现在以下几点:

  • 每个FFN层包含8个不同的"专家"(全连接层),根据门控值选取最优的2个进行激活
  • 输入序列中的每个token都会独立地选取专家,而不是整个序列对应一组专家
  • 实际参数量约为46.7B,在推理时激活的参数量约为13B

🚄 原生支持32K上下文(实测支持128K)

Mixtral模型原生支持32K上下文(实测可达128K)。用户可使用单一模型来解决不同长度的各类任务。

模型效果

  • 大模型竞技场:http://llm-arena.ymcui.com/
  • 生成效果:https://github.com/ymcui/Chinese-Mixtral#生成效果评测
  • 客观效果:https://github.com/ymcui/Chinese-Mixtral#客观效果评测

- Python
Published by ymcui about 2 years ago