vprak17/synthetic_audio_paired_preferences

一个合成的口语对话直接偏好优化（DPO）数据集，包含8482个示例，设计用于无文本语音语言模型的偏好对齐。每个示例对应一个双人对话中的单个助手轮次，提供口语提示（最后用户的发言，包括音频和文本）、选择的响应（模型生成的最佳口语回复，由LLM法官评分）和拒绝的响应（故意降级的回复，错误的响应类型或较低的LLM评分）。对话由Qwen3-32B合成生成，通过Voxtral-based TTS后端转换为语音，并由LLM法官评分。音频为24 kHz单声道WAV格式。

A synthetic spoken-dialogue DPO (Direct Preference Optimization) dataset with 8482 examples, designed for preference alignment of textless speech language models. Each example corresponds to a single assistant turn in a two-speaker conversation, providing the spoken prompt (the last user utterance, as audio + text), a chosen response (the best model-generated spoken reply scored by an LLM judge), and a rejected response (a deliberately degraded reply with wrong response type or lower LLM score). Conversations were synthetically generated with Qwen3-32B, converted to speech with a Voxtral-based TTS backend, and scored with an LLM judge. Audio is 24 kHz mono WAV.