MooreMuaMu/ancient-rl-grpo-pilot3-step100

Ancient Language GRPO — Pilot 3 Step 100

翻译任务最优 checkpoint：4 个翻译 bucket 全部涨点。

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 / 300
• 超参数: gen=16, temp=1.0, β=0.01, lr=2e-6

训练 Bucket (10 个)

与 pilot3-step50 相同的 10 个 bucket：

• 注释: tibetan, traditional_mongolian, manchu
• 断句: tangut, zhuang
• 修复: uyghur
• 翻译: classical_chinese, zhuang, traditional_mongolian, tangut

总训练数据: 60,606 samples

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

Bucket	Baseline	This Model	Δ
zhuang / segmentation	66.67	67.91	+1.24pp
classical_chinese / translation	42.04	42.68	+0.64pp
tangut / translation	42.09	42.63	+0.54pp
traditional_mongolian / annotation	57.04	58.43	+1.39pp
traditional_mongolian / translation	29.42	30.27	+0.85pp
uyghur / restoration	87.58	84.73	-2.85pp
tangut / segmentation	81.74	81.36	-0.38pp
manchu / annotation	65.44	63.93	-1.51pp
tibetan / annotation	57.73	56.65	-1.08pp
zhuang / translation	51.23	50.53	-0.70pp

最佳用途: 翻译类任务（文言文、传统蒙古文、西夏文翻译）

使用方式

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("MooreMuaMu/ancient-rl-grpo-pilot3-step100", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("MooreMuaMu/ancient-rl-grpo-pilot3-step100")