---
base_model:
- Qwen/Qwen3-0.6B
datasets:
- zenglhardt/convfill-dataset
language:
- en
license: apache-2.0
pipeline_tag: text-generation
---

# Model Card for Qwen-0.6B-ConvFill

This model card implements a finetune of [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) for the conversational infill task described in the [ConvFill paper](https://arxiv.org/abs/2511.07397).

## Model Details

This model should be used respecting the original license of the base model, Qwen/Qwen3-0.6B. The dataset that was used to finetune this model can be found [here](https://huggingface.co/datasets/zenglhardt/convfill-dataset).

### Model Description
Deploying responsive, multi-turn conversational voice agents with large language models poses a critical challenge: cloud-based foundation models utilize reasoning, information retrieval, and tool use for high-value tasks, but introduce latency that disrupts natural conversation. In contrast, small models can respond quickly but lack capabilities needed in real-world tasks. We propose _conversational infill_, a task where a small, local model generates prompt, contextually appropriate dialogue and seamlessly incorporates delayed, external knowledge produced in parallel by a foundation model backend. This finetune trains Qwen/Qwen3-0.6B to perform the conversational infill task.

- **Finetuned from model:** [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B)
- **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)

### Model Sources [optional]

- **Repository:** https://github.com/vysri/conversational-infill
- **Paper:** [Thinking While Speaking: Inference-Time Knowledge Transfer for Responsive and Intelligent Conversational Voice Agents](https://arxiv.org/abs/2511.07397)
- **Demo:** TBD

### Direct Use

This model is intended to be used with the infrastructure in the [ConvFill repository](https://github.com/vysri/conversational-infill).

## Bias, Risks, and Limitations

This model is not explicitly tuned for guardrailed behavior. Please use with caution.

## How to Get Started with the Model
Use the code in the [ConvFill repository](https://github.com/vysri/conversational-infill) to get started with this model.

### Training Data
A link to the training data for this model can be found [here](https://huggingface.co/datasets/zenglhardt/convfill-dataset). The dataset generation procedure can be found [here](https://github.com/zenglhardt/convfill-dataset). Information on training procedures can be found in the [ConvFill paper](https://arxiv.org/abs/2511.07397). Training code and scripts can be found in the [ConvFill repository](https://github.com/vysri/conversational-infill). 

## Citation
```
@misc{srinivas2026thinkingspeakinginferencetimeknowledge,
      title={Thinking While Speaking: Inference-Time Knowledge Transfer for Responsive and Intelligent Conversational Voice Agents}, 
      author={Vidya Srinivas and Zachary Englhardt and Vikram Iyer and Shwetak Patel},
      year={2026},
      eprint={2511.07397},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.07397}, 
}
```