Aller au contenu
ModelPrésentéExterne
Télécharger

SDiaReward 7B

SDiaReward-7B is a reward model for spoken dialogue, built on Qwen2.5-Omni-7B with a pooling layer and a linear scoring head (QwenOmniThinkerReward). Given a multi-turn conversation containing interleaved speech and...

Sourcehuggingface.co/MYJOKERML/SDiaReward-7B/tree/5119425ba2f77b2860998290d0d469b5c411b3ff
TypeOtherModèle de baseQwen/Qwen2.5-Omni-7BVersionRevision 5119425BClasse

Source du modèle

Type
Other
Taille du fichier
2,70–4,65 GB
Modèle de base
Qwen/Qwen2.5-Omni-7B
Version
vRevision 5119425
Mis à jour
12 oct. 2026
Téléchargements
12
Source
Hugging Face

Description de la source

SDiaReward-7B

SDiaReward-7B is a reward model for spoken dialogue, built on Qwen2.5-Omni-7B with a pooling layer and a linear scoring head (QwenOmniThinkerReward). Given a multi-turn conversation containing interleaved speech and text, it produces a scalar reward that reflects two qualities:

  • Modality-awareness — prosody, emotion, and acoustic naturalness (real human speech vs. synthetic TTS).

  • Colloquialness — spontaneous spoken style vs. scripted written style.

  • 📄 Paper: Modeling and Benchmarking Spoken Dialogue Rewards with Modality and Colloquialness (ACL 2026 Main Conference) — arXiv:2603.14889

  • 💻 Code: https://github.com/MM-Speech/SDiaReward

  • 📚 Training data (gated): SDiaReward dataset.

  • 🧩 Smaller variant: SDiaReward-3B.

Evaluation

On the ESDR-Bench validation set:

ModelAccuracyEval lossMargin
SDiaReward-7B0.9710.3581.167
SDiaReward-3B0.9170.4190.869

Usage

This checkpoint uses a custom reward architecture (QwenOmniThinkerReward: a Qwen2.5-Omni backbone with a pooling layer and a scalar reward head). The custom modeling code is not bundled in this repository, so the weights cannot be loaded with a plain AutoModel/pipeline call. To load the model and score conversations, follow the loading and inference instructions in the official code repository:

👉 https://github.com/MM-Speech/SDiaReward (checkpoint id: MYJOKERML/SDiaReward-7B)

Training

Trained with TRL's reward trainer on the SDiaReward preference dataset (11,630 episode-level preference pairs, ~200 hours of paired speech). Backbone: Qwen2.5-Omni-7B; pooling variant: mean_center; 2 epochs.

License & intended use

Released under Apache-2.0 for research use. The reward signal is intended for evaluating and improving spoken-dialogue systems; it is not a safety classifier.

Citation

@article{lu2026modeling,
  title={Modeling and benchmarking spoken dialogue rewards with modality and colloquialness},
  author={Lu, Jingyu and Wang, Yuhan and Zhuo, Fan and Cheng, Xize and Pan, Changhao and Pu, Xueyi and Chen, Yifu and Wen, Chenyuhao and Liang, Tianle and Zhao, Zhou},
  journal={arXiv preprint arXiv:2603.14889},
  year={2026}
}

Sources

1 source
Hugging FacePrincipalVersion 5119425ba2f77b2860998290d0d469b5c411b3ff
huggingface.co/MYJOKERML/SDiaReward-7B/tree/5119425ba2f77b2860998290d0d469b5c411b3ff

Vérifié 15 août

Artefacts du modèle

4 artefacts

model-00001-of-00004.safetensors

safetensors · 4,64 GB · SHA-256 d6b4b07b8181…d1b8 · Hugging Face

Télécharger

model-00002-of-00004.safetensors

safetensors · 4,65 GB · SHA-256 cb2c14b80bbc…c89e · Hugging Face

Télécharger

model-00003-of-00004.safetensors

safetensors · 4,65 GB · SHA-256 5bfc0ec435bc…4a05 · Hugging Face

Télécharger

model-00004-of-00004.safetensors

safetensors · 2,70 GB · SHA-256 904f04a577d6…8810 · Hugging Face

Télécharger

Extraits de sources

2 extraits
Readme15 aoûthuggingface.co/MYJOKERML/SDiaReward-7B

--- license: apache-2.0 language: - en base_model: - Qwen/Qwen2.5-Omni-7B pipeline_tag: audio-classification library_name: transformers tags: - reward-model - spoken-dialogue - speech - preference - qwen2_5_omni - modality-awareness - colloquialness --- # SDiaReward-7B **SDiaReward-7B** is a reward model for **spoken dialogue**, built on **Qwen2.5-Omni-7B** with a pooling layer and a linear scoring head (`QwenOmniThinkerReward`). Given a multi-turn conversation containing interleaved speech and text, it produces a scalar reward that reflects two qualities: - **Modality-awareness** — prosody, emotion, and acoustic naturalness (real human speech vs. synthetic TTS). - **Colloquialness** — spontaneous *spoken* style vs. scripted *written* style. - 📄 Paper: *Modeling and Benchmarking Spoken Dialogue Rewards with Modality and Colloquialness* (ACL 2026 Main Conference) — [arXiv:2603.14889](https://arxiv.org/abs/2603.14889) - 💻 Code: https://github.com/MM-Speech/SDiaReward - 📚 Training data (gated): [`SDiaReward` dataset](https://huggingface.co/datasets/MYJOKERML/SDiaReward). - 🧩 Smaller variant: [`SDiaReward-3B`](https://huggingface.co/MYJOKERML/SDiaReward-3B). ## Evaluation On the ESDR-Bench validation set: | Model | Accuracy | Eval loss | Margin | |---|---|---|---| | **SDiaReward-7B** | **0.971** | 0.358 | 1.167 | | SDiaReward-3B | 0.917 | 0.419 | 0.869 | ## Usage This checkpoint uses a **custom reward architecture** (`QwenOmniThinkerReward`: a Qwen2.5-Omni backbone with a pooling layer and a scalar reward head). The custom modeling code is **not** bundled in this repository, so the weights cannot be loaded with a plain `AutoModel`/`pipeline` call. To load the model and score conversations, follow the loading and inference instructions in the official code repository: 👉 **https://github.com/MM-Speech/SDiaReward** (checkpoint id: `MYJOKERML/SDiaReward-7B`) ## Training Trained with TRL's reward trainer on the [SDiaReward preference dataset](https://huggingface.co/datasets/MYJOKERML/SDiaReward) (11,630 episode-level preference pairs, ~200 hours of paired speech). Backbone: Qwen2.5-Omni-7B; pooling variant: `mean_center`; 2 epochs. ## License & intended use Released under Apache-2.0 for research use. The reward signal is intended for evaluating and improving spoken-dialogue systems; it is not a safety classifier. ## Citation ```bibtex @articl...

Note15 aoûthuggingface.co/MYJOKERML/SDiaReward-7B

Source context: 12 downloads · 0 likes · Pipeline audio-classification · Library transformers · Repo MYJOKERML/SDiaReward-7B