137 lines
4.0 KiB
Markdown
137 lines
4.0 KiB
Markdown
|
---
|
|||
|
|
language:
|
|||
|
|
- zh
|
|||
|
|
tags:
|
|||
|
|
- olmo3
|
|||
|
|
- chinese
|
|||
|
|
- sft
|
|||
|
|
- chat
|
|||
|
|
- fine-tuned
|
|||
|
|
- rtx4060
|
|||
|
|
license: apache-2.0
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
base_model: ynanxiu/olmo3-190M-zh-full
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# OLMo3-190M-zh-sft
|
|||
|
|
|
|||
|
|
**187M 参数中文 SFT 对话模型**——让预训练模型学会聊天。
|
|||
|
|
|
|||
|
|
基于 [OLMo3-190M-zh-full](https://huggingface.co/ynanxiu/olmo3-190M-zh-full) 进行有监督微调（SFT），专为中文对话场景训练。
|
|||
|
|
|
|||
|
|
## 📊 模型信息
|
|||
|
|
|
|||
|
|
| 项目 | 详情 |
|
|||
|
|
|------|------|
|
|||
|
|
| **基座模型** | [ynanxiu/olmo3-190M-zh-full](https://huggingface.co/ynanxiu/olmo3-190M-zh-full) |
|
|||
|
|
| **架构** | OLMo3, 12层, d=768, 12头, SwiGLU |
|
|||
|
|
| **参数量** | 187M |
|
|||
|
|
| **词表大小** | 48,000 |
|
|||
|
|
| **训练方式** | SFT（Supervised Fine-Tuning） |
|
|||
|
|
| **训练数据** | `cmz1024/llm101-olmo3-zh-demo-data` (SFT 对话格式) |
|
|||
|
|
| **格式** | ChatML (`<|user|>...<|assistant|>...`) |
|
|||
|
|
|
|||
|
|
## 🏋️ 训练详情
|
|||
|
|
|
|||
|
|
| 参数 | 值 |
|
|||
|
|
|------|-----|
|
|||
|
|
| 硬件 | NVIDIA RTX 4060 (8GB) |
|
|||
|
|
| 精度 | bfloat16 |
|
|||
|
|
| Batch Size | 1 × 32 (gradient_accumulation) |
|
|||
|
|
| 学习率 | 2e-5 |
|
|||
|
|
| LR Schedule | cosine + warmup (5%) |
|
|||
|
|
| Epochs | 2 |
|
|||
|
|
| 总步数 | 56,470 |
|
|||
|
|
| 训练耗时 | 31 小时 |
|
|||
|
|
|
|||
|
|
## 📈 训练指标
|
|||
|
|
|
|||
|
|
| 阶段 | Eval Loss | Accuracy | Entropy |
|
|||
|
|
|------|-----------|----------|---------|
|
|||
|
|
| 初始 (500步) | 3.36 | 43.1% | 3.388 |
|
|||
|
|
| 最终 (56,470步) | **2.24** | **54.6%** | **2.572** |
|
|||
|
|
|
|||
|
|
Loss 从 3.36 降至 2.24，下降 33%；accuracy 从 43% 升至 55%。
|
|||
|
|
|
|||
|
|
## 💬 使用方法
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|||
|
|
import torch
|
|||
|
|
|
|||
|
|
model_id = "ynanxiu/olmo3-190M-zh-sft"
|
|||
|
|
|
|||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
|
|||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|||
|
|
model_id,
|
|||
|
|
torch_dtype=torch.bfloat16,
|
|||
|
|
attn_implementation="sdpa",
|
|||
|
|
trust_remote_code=True,
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 对话模式（推荐）
|
|||
|
|
messages = [{"role": "user", "content": "你好，请用简单的语言解释什么是机器学习"}]
|
|||
|
|
formatted = tokenizer.apply_chat_template(
|
|||
|
|
messages, tokenize=False, add_generation_prompt=True
|
|||
|
|
)
|
|||
|
|
inputs = tokenizer(formatted, return_tensors="pt")
|
|||
|
|
|
|||
|
|
with torch.no_grad():
|
|||
|
|
outputs = model.generate(
|
|||
|
|
**inputs,
|
|||
|
|
max_new_tokens=200,
|
|||
|
|
temperature=0.3, # ⚠️ 推荐 0.3，避免重复
|
|||
|
|
top_p=0.9,
|
|||
|
|
do_sample=True,
|
|||
|
|
repetition_penalty=1.15,
|
|||
|
|
pad_token_id=tokenizer.pad_token_id,
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
response = tokenizer.decode(outputs[0][len(inputs["input_ids"][0]):], skip_special_tokens=True)
|
|||
|
|
print(response)
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### ⚙️ 推荐生成参数
|
|||
|
|
|
|||
|
|
| 参数 | 推荐值 | 说明 |
|
|||
|
|
|------|--------|------|
|
|||
|
|
| temperature | **0.3** | 190M 模型不建议超过 0.5，高温导致严重重复 |
|
|||
|
|
| repetition_penalty | 1.15~1.2 | 抑制重复 |
|
|||
|
|
| top_p | 0.9 | 标准核采样 |
|
|||
|
|
| max_new_tokens | 150~200 | 太长容易退化 |
|
|||
|
|
|
|||
|
|
## 🎮 在线体验
|
|||
|
|
|
|||
|
|
👉 **[HuggingFace Space](https://huggingface.co/spaces/ynanxiu/olmo3-190M-zh-sft)** — 直接在浏览器里对话
|
|||
|
|
|
|||
|
|
## 📝 生成效果
|
|||
|
|
|
|||
|
|
### 成功案例 ✅
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
用户：你好，请问你能帮我做什么？
|
|||
|
|
助手：当然可以！请告诉我你最近在做什么，或者有什么具体问题
|
|||
|
|
我可以帮你解决？我会尽力提供帮助。
|
|||
|
|
|
|||
|
|
用户：今天天气不错，我准备
|
|||
|
|
助手：我无法提供实时天气信息或天气预报。建议您查看天气应用
|
|||
|
|
或网站获取最新天气信息。如果您有其他问题，欢迎随时提问！
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### 已知限制 ⚠️
|
|||
|
|
|
|||
|
|
- **重复问题**：190M 参数较小，长回复可能出现重复。建议 temperature ≤ 0.3 + repetition_penalty ≥ 1.15。
|
|||
|
|
- **知识幻觉**：可能生成不准确的事实（如自称 "minimind"），SFT 数据影响了模型认知。
|
|||
|
|
- **复杂推理**：190M 参数量不适合复杂多轮推理任务。
|
|||
|
|
|
|||
|
|
## 🔗 相关模型
|
|||
|
|
|
|||
|
|
| 模型 | 类型 | 链接 |
|
|||
|
|
|------|------|------|
|
|||
|
|
| OLMo3-190M-zh-full | 基础预训练 | [HF](https://huggingface.co/ynanxiu/olmo3-190M-zh-full) |
|
|||
|
|
| OLMo3-190M-zh-nano | 小模型持续预训练 | [HF](https://huggingface.co/ynanxiu/olmo3-190M-zh-nano) |
|
|||
|
|
|
|||
|
|
## 📜 License
|
|||
|
|
|
|||
|
|
Apache 2.0
|