Qwen3.5-27B Ancient RL R32 Step 250 LoRA

LoRA adapter checkpoint from the Ancient Swift RL Stage 2 GRPO run.

This repo contains adapter files only, not the dense base model and not optimizer states. Use it with the matching base model: qwen3.5-27b-ancient-stage2-checkpoint-227-merged.

Selection status

recommended checkpoint; statistically significant held-out improvement

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

Run root: /blob/experiments/ancient-swift-rl-stage2/qwen35_27b_ckpt227_uy_mn_bo_ta_full_r32_lr1e4/v0-20260811-070520
Source checkpoint: checkpoint-250
LoRA rank/alpha: 32/32
LoRA target modules: all linear
Dropout: 0.05
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.

Files

adapter_config.json: 1501f59a7650bc578076e944a67f8f49ed7a63d2ea15e12a7ac5d5a4b365d21f
adapter_model.bin: 9fb15ba4de628dece1afe8dd2df3f3f6cb81ef356dbbf8e9e402a50875765b16
additional_config.json: c7799462ebedae6557ffad31566029e2d2f958b7b40e46e972cf901bcaf45733
args.json: 585d7eadc91949ed705bb8beec00fd031fe748b18e539bd4ffa08265c2d150e0
trainer_state.json: 62f13c62c729907fd905c8ad7fcb71c2d68a1025313e04d8217c56bdb482cf45
training_args.bin: b5acf749719b965b3fd31820c768af15a06d8d344aee4b4165924184407ce2ea

Notes

This is an experimental LoRA adapter. For production use, run the full canonical held-out evaluation before deployment.

Downloads last month: 6