qwen35-27b-ancient-rl-r32-step200
This repository contains the full merged bfloat16 model for the rank-32 GRPO checkpoint at step 200. 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-step200-lora.
Intended Use
from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "MooreMuaMu/qwen35-27b-ancient-rl-r32-step200"
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-200 - 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 200 | Delta | 95% CI |
|---|---|---|---|---|
| Exact | 0.3367 | 0.3100 | -0.0267 | [-0.0700, +0.0133] |
| Char-F1 | 0.5708 | 0.5758 | +0.0050 | [-0.0320, +0.0431] |
| BERTScore-F1 | 0.7627 | 0.7901 | +0.0274 | [-0.0010, +0.0571] |
| Semantic composite | 0.6061 | 0.6139 | +0.0078 | [-0.0255, +0.0415] |
| Has answer | 0.9400 | 0.9867 | +0.0467 | [+0.0167, +0.0767] |
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 200 | Delta |
|---|---|---|---|
| Corpus BLEU-2 (%) | 34.59 | 37.80 | +3.21 pp |
| Sentence BLEU-2 mean (%) | 48.42 | 47.59 | -0.82 pp |
| Nonempty extracted answer | 93.7% | 98.7% | +5.0 pp |
Bucket-level SacreBLEU2:
| Bucket | Corpus BLEU-2 base -> step200 | Delta | Sentence BLEU-2 base -> step200 | Delta | Step200 nonempty |
|---|---|---|---|---|---|
| Tibetan annotation | 65.98 -> 59.98 | -6.00 pp | 74.86 -> 66.85 | -8.01 pp | 96.0% |
| Tibetan translation | 14.16 -> 21.22 | +7.06 pp | 14.20 -> 18.87 | +4.67 pp | 100.0% |
| Traditional Mongolian annotation | 48.69 -> 54.63 | +5.94 pp | 48.83 -> 47.03 | -1.79 pp | 98.0% |
| Traditional Mongolian translation | 33.79 -> 30.38 | -3.41 pp | 31.77 -> 32.50 | +0.73 pp | 100.0% |
| Uyghur annotation | 81.43 -> 89.15 | +7.72 pp | 85.26 -> 89.94 | +4.68 pp | 98.0% |
| Uyghur translation | 39.48 -> 30.69 | -8.79 pp | 35.57 -> 30.35 | -5.21 pp | 100.0% |
Decision note: step 200 is an exploratory checkpoint. It improves answer presence significantly and has positive BERTScore/SacreBLEU2 corpus movement, but aggregate semantic confidence intervals still cross zero and translation buckets are mixed.