ConvFill Dataset
收藏资源简介:
ConvFill数据集是一个包含290,571个训练示例的语料库,用于对话填充任务:通过将运行在设备上的小型低延迟说话者模型与在后台推理、检索和调用工具的大型云端推理者模型配对,以保持语音代理的响应性。说话者立即响应用户,并在推理者的知识流式传输时融入其中,从而确保对话在推理者工作时不会停滞。
The ConvFill dataset is a corpus containing 290,571 training examples for the dialogue filling task. It pairs on-device small low-latency speaker models with large cloud-based reasoner models that perform background inference, retrieval and tool invocation to maintain the responsiveness of voice agents. The speaker responds to users instantly, and incorporates the streamed knowledge outputs from the reasoner into its responses, thus ensuring that the conversation does not stall while the reasoner is working.
ConvFill 数据集概述
ConvFill 数据集是一个用于**对话填充(Conversational Infill)**任务的语料库,包含 290,571 个训练样本。该任务旨在通过一个本地运行的轻量级 Talker 模型 与一个云端大模型的 Reasoner 模型 协作,使得语音智能体在 Reasoner 进行推理、检索或调用工具的同时,Talker 能够即时回应用户,避免对话中断。
数据规模与组成
数据集包含 8,443 个对话,共计 74,508 个轮次,每个轮次包含多个短语级别的训练样本。具体数据文件及统计如下:
| 文件 | 对话数 | 轮次数 | 短语样本数 | <sil> 填充样本数 |
知识样本数 |
|---|---|---|---|---|---|
conversations_advice.jsonl |
1,000 | 8,308 | 36,710 | 12,342 | 24,368 |
conversations_assistant.jsonl |
1,000 | 8,408 | 37,495 | 12,676 | 24,819 |
conversations_support.jsonl |
1,000 | 8,313 | 35,862 | 12,147 | 23,715 |
conversations_education.jsonl |
1,000 | 8,421 | 37,153 | 12,481 | 24,672 |
conversations_medical.jsonl |
1,000 | 8,400 | 37,246 | 12,540 | 24,706 |
conversations_planning.jsonl |
1,005 | 8,474 | 37,488 | 12,655 | 24,833 |
conversations_dstc8_scaffold.jsonl |
2,438 | 24,184 | 68,617 | 36,711 | 31,906 |
| 总计 | 8,443 | 74,508 | 290,571 | 111,552 | 179,019 |
数据全部为英文。前六个文件覆盖了建议、通用助手查询、客服、教育、医疗和规划领域。conversations_dstc8_scaffold.jsonl 基于 Schema-Guided Dialogue (SGD / DSTC8) 对话框架生成。
数据格式
每个 JSONL 记录包含一个顶层 conversation 数组,每个元素代表一个对话轮次,结构如下:
user: 当前轮次的用户话语。thoughts: Reasoner 模型生成的思考流,每个元素是一条知识片段或<sil>(表示沉默)标记。response: Talker 模型输出的响应流,与thoughts一一对应。len(thoughts) == len(response):保证输入输出序列长度相等。response[i]与thoughts[i]对齐:若thoughts[i]为<sil>,则response[i]为不依赖未获取知识的填充语;否则,response[i]应对话式地传达thoughts[i]的知识。
示例 JSON 结构: json { "conversation": [ { "user": "Ive been having a really hard time falling asleep lately...", "thoughts": ["<sil>", "How long has this been going on?", "Does this happen at the same time every night or does it vary?"], "response": ["Oh man, that sounds rough.", "How long would you say this has been going on?", "And is it the same time every night, or does it vary a lot?"] } ] }
数据加载
数据集可通过 Python 标准库加载,每个 (thoughts[i], response[i]) 对通常被转换为一个短语级别的训练样本,上下文包含之前的对话轮次及当前轮次中先前的短语。
仓库结构
data/: 已发布的 JSONL 数据集与统计数据。generated_data/: 生成管道的默认输出目录。configs/: 各子集的生成配置文件。examples/: 生成提示中使用的少样本示例。topics/: 自由生成子集的主题种子。prompt_template.txt和prompt_template_scaffold.txt: 生成提示模板。src/pipeline/: 数据集生成与 LLM 客户端代码。src/validators/: 结构、NLI、BERTScore 和骨架验证。src/evals/generated_data_stats.py: 重新生成dataset_stats.csv。
许可证
- 代码与支持文件(
src/、configs/等)采用 MIT 许可证。 - 数据(
data/、topics/、examples/)采用 CC BY-SA 4.0 许可证。 - DSTC8/SGD 骨架子集源自 Schema-Guided Dialogue 数据集,亦采用 CC BY-SA 4.0。
引用
若使用该数据集,请引用以下论文:
bibtex @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}, }





