MagicLuke/qwen25omni-sicl-grpo-best

Qwen2.5-Omni SICL GRPO Best

This repository contains a PEFT LoRA adapter for Qwen/Qwen2.5-Omni-7B. It is the best currently selected Qwen GRPO checkpoint from the SICL-GRPO experiments, selected by RSR 3-shot WER.

Checkpoint

• Source checkpoint: checkpoint-600
• Training job: 18288799
• Recipe: ctx_good_perf_bad_2k, seed 42
• Reward setting: ICE-sim soft reward weight w=0.04
• Local source path at export time: checkpoint/Qwen2.5-Omni-7B/lora_grpo_ctx_good_perf_bad_2k_s42_ctx0p6_wer0p1_ice_soft_w004/18288799/v0-20260516-181131/checkpoint-600

Evaluation

ASR results report corpus-level WER. Bounded WER caps each utterance contribution at 1.0 before aggregation.

Eval	Shot	Raw WER	Bounded WER	Accuracy
RSR selected eval	3	16.79%	16.56%
RSR full eval	0	29.15%	28.73%
RSR full eval	3	16.63%	16.32%
MYST full eval	0	16.95%	16.94%
MYST full eval	3	11.25%	11.23%
MMAR full eval	0			49.50%
MMAR full eval	3			56.40%
MMAU full eval	0			66.80%
MMAU full eval	3			72.80%

Usage

Load this adapter with PEFT on top of Qwen/Qwen2.5-Omni-7B using the same inference stack as the SICL-GRPO experiments.

from peft import PeftModel
from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor

base = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-Omni-7B",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, "MagicLuke/qwen25omni-sicl-grpo-best")
processor = Qwen2_5OmniProcessor.from_pretrained(
    "Qwen/Qwen2.5-Omni-7B",
    trust_remote_code=True,
)

For audio inference, use the project inference scripts and message formatting for Qwen2.5-Omni.

Files

This upload is adapter-only. It intentionally excludes optimizer, scheduler, RNG, and trainer state.