Open-weight model
jevtrainer-omni3b-2026-10-04
by Weihongliang weihongliang/jevtrainer-omni3b-2026-10-04
jevtrainer-omni3b-2026-10-04 is an open-weight model from Weihongliang, released under Apache License 2.0. Its published files total 1.9 GB. It draws 18 downloads a month.
A Jev-style decision model that reads text, images, video, and audio. The readout is marker. LoRA (r=32, alpha=64) is trained on the thinker of Qwen/Qwen2.5-Omni-3B. Only the thinker is loaded; the speech output is not. Training finished on 2026-10-04.
Model Card
By Weihongliang, published under apache-2.0, revision 0e4aa2644994.
A Jev-style decision model that reads text, images, video, and audio. The readout is marker. LoRA (r=32, alpha=64) is trained on the thinker of Qwen/Qwen2.5-Omni-3B. Only the thinker is loaded; the speech output is not. Training finished on 2026-10-04. The training config is configs/av/omni3bv4.yaml in jevtrainer: 79 text, image, audio, and video datasets, one epoch. The holdout has 1,610 examples, accuracy 85.03%, skill 79.54%, ECE 0.0096. Accuracy in percent. This repo is the LoRA adapter and the readout. Base weights still come from Qwen/Qwen2.5-Omni-3B. readout.json records the base model id.
Read Weihongliang's full model card
A Jev-style decision model that reads text, images, video, and audio. The readout is marker. LoRA (r=32, alpha=64) is trained on the thinker of Qwen/Qwen2.5-Omni-3B. Only the thinker is loaded; the speech output is not. Training finished on 2026-10-04.
The training config is configs/av/omni3b_v4.yaml in jevtrainer: 79 text, image, audio, and video datasets, one epoch. The holdout has 1,610 examples, accuracy 85.03%, skill 79.54%, ECE 0.0096.
Evaluation
Accuracy in percent.
| Model | WorldSense | OmniBench | Daily-Omni | AV-Odyssey | SpeakerBench |
|---|---|---|---|---|---|
| Qwen2.5-Omni-3B | 34.87 | 43.82 | 49.79 | 31.99 | 39.69 |
| JevTrainer Omni 3B | 37.05 | 49.43 | 54.05 | 36.60 | 43.69 |
pip install -e . # github.com/hongliang-wei/jevtrainer
huggingface-cli download weihongliang/jevtrainer-omni3b-2026-10-04 --local-dir runs/jev-omni3b
jt eval configs/eval/av_omni.yaml --set checkpoint=runs/jev-omni3b --set benchmarks=worldsense,omnibench
This repo is the LoRA adapter and the readout. Base weights still come from Qwen/Qwen2.5-Omni-3B. readout.json records the base model id.
Identity and Version
- Repository
- weihongliang/jevtrainer-omni3b-2026-10-04
- Publisher
- Weihongliang
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- Not stated by the source
- Languages
- jev
- Revision
- 0e4aa26449948943e11ad6c30798558958d95ff8
- First published
- 2026-10-04
- Last updated
- 2026-10-09
Files and Weights
26 files, 1.9 GB in total. The weights are 12 files totalling 1.9 GB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| adapter/adapter_model.safetensors | Weights | 239.5 MB | ae153d50075e |
| model/model-00006-of-00039.safetensors | Weights | 31.5 MB | 74753ac12e57 |
| model/model-00009-of-00039.safetensors | Weights | 218.1 MB | 271605763f05 |
| model/model-00010-of-00039.safetensors | Weights | 218.1 MB | 69d2212ac5e5 |
| model/model-00012-of-00039.safetensors | Weights | 218.1 MB | 112f2957efba |
| model/model-00014-of-00039.safetensors | Weights | 218.1 MB | 0859190c6b95 |
| model/model-00021-of-00039.safetensors | Weights | 24.0 MB | b4bd4e1f2285 |
| model/model-00025-of-00039.safetensors | Weights | 218.1 MB | 370521630e58 |
| model/model-00027-of-00039.safetensors | Weights | 218.1 MB | 6fd726515f09 |
| model/model-00029-of-00039.safetensors | Weights | 218.1 MB | c96a97153140 |
| model/model-00039-of-00039.safetensors | Weights | 64.0 MB | 8d309cd54df0 |
| readout.safetensors | Weights | 8.3 KB | 7f1bc4c5a588 |
| adapter/adapter_config.json | Configuration | 1.3 KB | — |
| calibration.json | Configuration | 95 B | — |
| config.yaml | Configuration | 10.5 KB | — |
| metrics.json | Configuration | 795 B | — |
| model/config.json | Configuration | 4.8 KB | — |
| model/model.safetensors.index.json | Configuration | 110.4 KB | — |
| readout.json | Configuration | 435 B | — |
| tokenizer/processor_config.json | Configuration | 2.6 KB | — |
| README.md | Documentation | 1.7 KB | — |
| adapter/README.md | Documentation | 5.2 KB | — |
| tokenizer/chat_template.jinja | Other | 1.3 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer/tokenizer.json | Tokenizer | 11.4 MB | d7cdeec9a6f2 |
| tokenizer/tokenizer_config.json | Tokenizer | 1.0 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 1.9 GB
Released by Weihongliang through its official repository on Hugging Face. Read the license.
Built From
- Derived from Qwen/Qwen2.5-Omni-3B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 1.9 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About jevtrainer-omni3b-2026-10-04
Can I use jevtrainer-omni3b-2026-10-04 commercially?
Yes. jevtrainer-omni3b-2026-10-04 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.