---
language:
- zh
- mn
- ug
- bo
tags:
- ancient-languages
- grpo
- rl
- qwen3
license: apache-2.0
base_model: zisuh/round2-10lang-passk-15k-sft-0608
---

# Ancient Language GRPO — Pilot 2 Step 100

**注释与断句任务最强 checkpoint：tibetan_annotation +2.69pp。**

## 模型信息

- **基座模型**: Qwen3-14B (zisuh/round2-10lang-passk-15k-sft-0608 checkpoint-195-merged)
- **训练方法**: GRPO + LoRA rank=64
- **Reward**: BLEU-2 + BERTScore (batched)
- **训练步数**: 100 / 200
- **超参数**: gen=16, temp=1.0, **β=0.02**, lr=2e-6

## 训练 Bucket (6 个)

| Bucket | 任务 | 训练样本数 | test_final_new Δ |
|---|---|---|---|
| tibetan\_\_annotation | 注释 | 5,273 | **+2.69pp** ✅ |
| traditional\_mongolian\_\_annotation | 注释 | 2,403 | +0.36pp |
| xishuangbanna\_dai\_\_annotation | 注释 | 1,823 | -1.33pp |
| tangut\_\_segmentation | 断句 | 3,305 | +0.25pp |
| zhuang\_\_segmentation | 断句 | 222 | **+1.24pp** ✅ |
| uyghur\_\_restoration | 修复 | 4,335 | **+0.98pp** ✅ |

**总训练数据**: 30,113 samples

## 评测结果 (sacrebleu sentence-BLEU-2, 0-100)

| Bucket | Baseline | This Model | Δ |
|---|---|---|---|
| **tibetan / annotation** | 61.09 | **63.78** | **+2.69pp** ✅ |
| zhuang / segmentation | 66.67 | **67.91** | **+1.24pp** ✅ |
| uyghur / restoration | 89.04 | **90.02** | **+0.98pp** ✅ |
| uyghur / translation | 31.88 | **32.22** | +0.34pp |
| traditional\_mongolian / annotation | 59.65 | 60.01 | +0.36pp |
| tangut / segmentation | 84.72 | 84.97 | +0.25pp |

**最佳用途**: 注释和断句任务（尤其是藏文注释）

## 与 Pilot 3 的区别

| 参数 | Pilot 2 (本模型) | Pilot 3 |
|---|---|---|
| β (KL penalty) | **0.02** (更保守) | 0.01 (更探索) |
| Bucket 数 | 6 | 10 |
| 翻译 bucket | 无 | 4 个 |
| 最大单 bucket 提升 | **+2.69pp** | +1.54pp |
| 特点 | annotation/segmentation 专精 | 覆盖广、translation 涨点 |

β=0.02 对注释/断句/修复类任务更有效（这些任务需要保守探索），而翻译任务需要 β=0.01 的更激进探索。

## 使用方式

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("MooreMuaMu/ancient-rl-grpo-pilot2-step100", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("MooreMuaMu/ancient-rl-grpo-pilot2-step100")
```