README.md

# Posttrain Lab Weights

《大语言模型后训练实践》课程实验的 LoRA adapter 权重。

| 目录 | 说明 |
|------|------|
| `lab1-sft-r8/` | Lab1: LoRA r=8 SFT on Alpaca-zh (Qwen3-1.7B-Base) |
| `lab1-sft-r32/` | Lab1: LoRA r=32 SFT on Alpaca-zh (Qwen3-1.7B-Base) |
| `lab2-sft-best/` | Lab2: LoRA r=32 SFT on COIG-CQIA/zhihu (Qwen3-1.7B-Base) |
| `lab2-ablation-raw/` | Lab2 ablation: raw data (no QC) |
| `lab2-ablation-dedup/` | Lab2 ablation: dedup only |
| `lab2-ablation-clean/` | Lab2 ablation: full QC (dedup + filter) |
| `lab3-dpo/` | Lab3: DPO aligned on UltraFeedback (Qwen3-1.7B, beta=0.1) |
| `lab3-simpo/` | Lab3: SimPO aligned on UltraFeedback (Qwen3-1.7B, beta=2.0, gamma=0.5) |
| `lab3-dpo-beta005/` | Lab3 ablation: DPO beta=0.05 |
| `lab3-dpo-beta01/` | Lab3 ablation: DPO beta=0.1 |
| `lab3-dpo-beta05/` | Lab3 ablation: DPO beta=0.5 |

## 使用方法

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B-Base")
model = PeftModel.from_pretrained(model, "leixinlin/posttrain-lab-weights", subfolder="lab2-sft-best")
```