From 79a616ac157559d1c4d4b5c661287505b4fbf81e Mon Sep 17 00:00:00 2001 From: gongjy <2474590974@qq.com> Date: Wed, 28 Aug 2024 18:03:24 +0800 Subject: [PATCH] update readme's error --- README.md | 4 ++-- README_en.md | 6 +++--- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/README.md b/README.md index 9c10ef7..b11d38e 100644 --- a/README.md +++ b/README.md @@ -38,7 +38,7 @@ 因此，本项目的目标是把上手LLM的门槛无限降低， 直接从0开始训练一个极其轻量的语言模型。 -（截至2024.8.27）MiniMind首发包含3个型号模型，最小仅需26M（0.02B），即可具备Amazing的对话能力！ +（截至2024.8.28）MiniMind首发包含4个型号模型，最小仅需26M（0.02B），即可具备Amazing的对话能力！ | 模型 (大小) | 速度 (Tokens/s) | 推理占用 | 训练占用(`batch_size=8`) | |------------------------|---------------|--------|----------------------| @@ -175,7 +175,7 @@ python 2-eval.py 但MiniMind这里选择了mistral tokenizer作为分词器以保持整体参数轻量，避免头重脚轻，因为mistral的词表大小只有32,000。 且MiniMind在实际测试中几乎没有出现过生僻词汇解码失败的情况，效果良好。 - > 方便对比测试效果，额外训练了一个自定义Tokenizer模型的版本**MiniMind(-T)**，自定义词表压缩长度到6400，使得LLM总参数进一步降低到40M左右。 + > 方便对比测试效果，额外训练了一个自定义Tokenizer模型的版本**MiniMind-small-T**，自定义词表压缩长度到6400，使得LLM总参数进一步降低到26M左右。 --- diff --git a/README_en.md b/README_en.md index 7f2d0f7..5371255 100644 --- a/README_en.md +++ b/README_en.md @@ -45,7 +45,7 @@ exacerbates the problem of finding quality content to understand LLMs, severely Therefore, the goal of this project is to lower the barrier to entry for working with LLMs as much as possible, by training an extremely lightweight language model from scratch. -(As of August 27, 2024) The initial release of MiniMind includes three model variants, with the smallest being just +(As of August 28, 2024) The initial release of MiniMind includes four model variants, with the smallest being just 26MB (0.02B) and still exhibiting amazing conversational capabilities! | Model (Size) | Speed (Tokens/s) | Inference Memory | Training Memory (`batch_size=8`) | @@ -73,7 +73,7 @@ We hope this open-source project helps LLM beginners get started quickly! 👉**Recent Updates**
- 2024-08-27 + 2024-08-28 - Project first open-sourced
@@ -206,7 +206,7 @@ git clone https://github.com/jingyaogong/minimind.git performance in practical tests, with almost no failures in decoding rare words. > For comparison purposes, an additional custom Tokenizer version **MiniMind(-T)** was trained, reducing the - vocabulary size to 6,400, which further decreases the total model parameters to around 40M. + vocabulary size to 6,400, which further decreases the total model parameters to around 26M. ---