This repository contains full merged model weights. It is not a LoRA-only adapter repo. Users can load this repo directly with a Transformers version that supports qwen3_5 / Qwen3_5ForConditionalGeneration.

The corresponding adapter-only repo is MooreMuaMu/qwen35-27b-ancient-rl-r32-step250-lora.

Usage

python

import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

repo_id = "MooreMuaMu/qwen35-27b-ancient-rl-r32-step250"
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    repo_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(repo_id)

For text-only serving, use the same chat template/tokenizer files included in this repo.

Selection status

Recommended checkpoint from the Ancient Swift RL Stage 2 R32 run. In the held-out 300-prompt small eval, step250 was the best checkpoint and showed statistically significant improvement versus the base model.

Do not use later checkpoints from this run (step300, step350, step400, step450) for model selection: held-out quality degraded at step300 and collapsed after step350.

Training run

Base model: /blob/models/qwen35-27b-ancient-stage2/checkpoint-227-merged
Source LoRA checkpoint: /blob/experiments/ancient-swift-rl-stage2/qwen35_27b_ckpt227_uy_mn_bo_ta_full_r32_lr1e4/v0-20260811-070520/checkpoint-250
LoRA rank/alpha: 32/32
LoRA dropout: 0.05
LoRA target modules: all linear target modules in the Qwen3.5 language model
RL algorithm: GRPO
LR: 1e-4
GRPO beta: 0.001
Training data: Uyghur, traditional Mongolian, Tibetan translation + annotation

Held-out small eval versus base

300 prompts, 50 per bucket, thinking on, temperature 0.7, top-p 0.95, n=1, seed 42. Prompt-level paired bootstrap: 10,000 samples, seed 20260811.

Metric	Delta vs base
Exact	+0.0400 CI [+0.0033,+0.0800]
Char-F1	+0.0539 CI [+0.0160,+0.0910]
BERTScore-F1	+0.0631 CI [+0.0330,+0.0942]
Semantic composite	+0.0555 CI [+0.0222,+0.0880]
Has ANS	+0.0667 CI [+0.0333,+0.1000]
Adapter has_ans	0.9867
SacreBLEU-2 corpus delta	+7.98 pp
SacreBLEU-2 sentence mean delta	+4.21 pp

The SacreBLEU-2 numbers use sacrebleu.metrics.bleu.BLEU(tokenize="zh", max_ngram_order=2, smooth_method="exp", effective_order=True) after extracting <ANS>...</ANS> from the completion.

Weight files

Total parameters in index metadata: 27,356,728,560
Total tensor bytes in index metadata: 54,713,457,120
Safetensors shards: 12
model-00001-of-00012.safetensors: 2,542,796,928 bytes
model-00002-of-00012.safetensors: 4,842,451,920 bytes
model-00003-of-00012.safetensors: 4,965,227,944 bytes
model-00004-of-00012.safetensors: 4,912,819,264 bytes
model-00005-of-00012.safetensors: 4,986,198,544 bytes
model-00006-of-00012.safetensors: 4,912,819,320 bytes
model-00007-of-00012.safetensors: 4,932,703,272 bytes
model-00008-of-00012.safetensors: 4,966,314,576 bytes
model-00009-of-00012.safetensors: 4,964,162,248 bytes
model-00010-of-00012.safetensors: 4,933,789,824 bytes
model-00011-of-00012.safetensors: 4,965,228,032 bytes
model-00012-of-00012.safetensors: 2,789,094,896 bytes

Notes

This is an experimental merged checkpoint. Run your own task-specific evaluation before production use.