qwen35-27b-ancient-rl-r32-step100

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:

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>.

Bucket-level SacreBLEU2:

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.