README.md

---
license: apache-2.0
language:
- zh
tags:
- llm001
- olmo3
- chinese
- pretrained
---

# OLMo3-190M-zh-full

为零基础 AI 大模型研发训练营（llm001）L04 基础预训练模型（190M 参数，2000 步训练）。

## 模型配置

- hidden_size: 768, num_layers: 12, num_heads: 12, intermediate_size: 3072
- vocab_size: 48000, sliding_window: 4096

## 训练配置

- 数据：cmz1024/llm101-olmo3-zh-demo-data (500M tokens)
- 训练：A100, max_steps=2000, bs=12×10=120, lr=5e-4, bf16
- Loss: ~9 → ~4.7, 平均 loss 5.14

## 用法

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("jinxing6609/olmo3-190m-zh-full")
tok = AutoTokenizer.from_pretrained("jinxing6609/olmo3-190m-zh-full")
```