qwen35-27b-ancient-rl-r32-step100
This repository contains the full merged bfloat16 model for the rank-32 GRPO checkpoint at step 100. It is not an adapter-only upload: the LoRA adapter has been merged into the Qwen3.5-27B ancient-language stage-2 base model, and the repository includes the safetensors shards plus tokenizer/processor/chat-template files needed for direct transformers loading.
Adapter-only sibling: MooreMuaMu/qwen35-27b-ancient-rl-r32-step100-lora.
Intended Use
from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "MooreMuaMu/qwen35-27b-ancient-rl-r32-step100"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
repo_id,
torch_dtype="auto",
device_map="auto",
trust_remote_code=True,
)
Training Source
- Base model:
/blob/models/qwen35-27b-ancient-stage2/checkpoint-227-merged - LoRA checkpoint:
/blob/experiments/ancient-swift-rl-stage2/qwen35_27b_ckpt227_uy_mn_bo_ta_full_r32_lr1e4/v0-20260811-070520/checkpoint-100 - LoRA rank/alpha: 32/32
- Merge dtype: bfloat16
- Serialization: safetensors, 5GB max shard size
Evaluation Summary
Evaluation used the independent six-bucket held-out subset: 300 prompts, 50 per bucket, thinking enabled, temperature 0.7, top-p 0.95, n=1, seed 42. Deltas are prompt-paired against the same base-model generations where applicable. Bootstrap confidence intervals use 10,000 paired samples.
Overall semantic metrics:
| Metric | Base | Step 100 | Delta | 95% CI |
|---|---|---|---|---|
| Exact | 0.3133 | 0.3400 | +0.0267 | [-0.0167, +0.0733] |
| Char-F1 | 0.5499 | 0.5579 | +0.0080 | [-0.0328, +0.0495] |
| BERTScore-F1 | 0.7638 | 0.7657 | +0.0019 | [-0.0309, +0.0348] |
| Semantic composite | 0.5915 | 0.6015 | +0.0100 | [-0.0260, +0.0473] |
| Has answer | 0.9533 | 0.9467 | -0.0067 | [-0.0400, +0.0267] |
SacreBLEU2 answer-extraction metrics use sacrebleu.metrics.BLEU(tokenize="zh", max_ngram_order=2, smooth_method="exp", effective_order=True) after extracting text inside <ANS>...</ANS> and discarding any prefix before </think>.
| Metric | Base | Step 100 | Delta |
|---|---|---|---|
| Corpus BLEU-2 (%) | 34.30 | 39.86 | +5.56 pp |
| Sentence BLEU-2 mean (%) | 46.12 | 48.10 | +1.98 pp |
| Nonempty extracted answer | 95.0% | 94.7% | -0.3 pp |
Bucket-level SacreBLEU2:
| Bucket | Corpus BLEU-2 base -> step100 | Delta | Sentence BLEU-2 base -> step100 | Delta | Step100 nonempty |
|---|---|---|---|---|---|
| Tibetan annotation | 59.14 -> 56.73 | -2.41 pp | 65.40 -> 66.82 | +1.42 pp | 94.0% |
| Tibetan translation | 15.68 -> 18.15 | +2.47 pp | 16.43 -> 16.00 | -0.43 pp | 94.0% |
| Traditional Mongolian annotation | 53.65 -> 70.42 | +16.77 pp | 50.84 -> 56.84 | +6.01 pp | 92.0% |
| Traditional Mongolian translation | 25.52 -> 26.15 | +0.63 pp | 24.41 -> 29.26 | +4.85 pp | 100.0% |
| Uyghur annotation | 85.13 -> 90.75 | +5.62 pp | 87.89 -> 87.26 | -0.63 pp | 88.0% |
| Uyghur translation | 34.55 -> 36.67 | +2.12 pp | 31.77 -> 32.41 | +0.64 pp | 100.0% |
Decision note: step 100 is an exploratory checkpoint. Its aggregate semantic improvement is small and the confidence interval crosses zero, although SacreBLEU2 corpus score improves, mainly from Traditional Mongolian annotation.