Downloads · 30 days
11
22% of all-time downloads
qwen25omni-sicl-grpo-best
MagicLuke/qwen25omni-sicl-grpo-best
qwen25omni-sicl-grpo-best is a automatic speech recognition model from MagicLuke. Use it when you need speech turned into text. It is set up for peft.
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.
All-time downloads
49
Public
Repo size
40.4 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors40.4 MB · 100%
At a glance
- Task
- Automatic Speech Recognition
- Library
- peft
- Access
- Public
- Created
- May 18, 2026
- Updated
- May 18, 2026
- SHA
- 1f3d8dcc
Base models
Model card
From the Hugging Face model README
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.
- Task
- Automatic Speech Recognition
- Library
- peft
- Created
- May 18, 2026
- Updated
- May 18, 2026