yqyao/EgoMem-eval

EgoMem Evaluation Data

Dataset Description

This repository contains the evaluation data for EgoMem: Lifelong Memory Agent for Full-duplex Omnimodal Models.

The dataset is divided into two splits:

• original: contains the original user audio without added noise.
• with_noise: contains noisy versions of the user audio.

The two splits share the same data structure. The primary difference between them is the content of the user_wav field.

Data Fields

Each sample contains the following top-level fields:

Field	Description
meta	A unique identifier for the sample.
user_wav	A one-dimensional NumPy array containing the waveform of the user's input audio.
bot_wav	A one-dimensional NumPy array containing the ground-truth waveform of the model's response.
sampling_rate	The sampling rate of both user_wav and bot_wav.
dialog_records	A list containing multiple dialogue sessions that occur within user_wav and bot_wav. Each session is represented by a dictionary containing all annotations associated with that session.

Dialogue Session Structure

Each element in dialog_records represents one dialogue session and contains the following fields:

Field	Description
dialog_span	The time span of the session within user_wav and bot_wav, expressed in audio sample points.
dialog_span_in_seconds	The same session span expressed in seconds.
profile_summary	A structured summary containing the user's name, relevant facts, personal profile, and a summary of previous conversations.
relation_graph	The user's relation graph. It is represented as a list in which each element describes another person related to the current user, including the person's name, relationship to the user, and known facts.
dialog_turn_contents	A list containing the user's questions and the bot's ground-truth responses, together with their timestamps, text, retrieval annotations, and supporting facts.

profile_summary

The profile_summary dictionary contains the following fields:

Field	Description
person	The name of the current user.
person_fact	Facts associated with the current user.
person_profile	A general profile of the current user.
person_summary	A summary of the user's previous dialogue history.

relation_graph

Each element in relation_graph contains the following fields:

Field	Description
person	The name of a person related to the current user.
person_relation	The relationship between this person and the current user.
person_fact	A list of known facts about this person.

dialog_turn_contents

Each element in dialog_turn_contents describes one dialogue turn and contains information about the user's question and the bot's response.

Field	Description
user_stamp	The start timestamp of the user's question.
user_txt	The transcription of the user's question.
bot_stamp	The start timestamp of the bot's response.
bot_stamp_end	The end timestamp of the bot's response.
bot_stamp_end_audio	The end timestamp of the bot's audio response.
bot_txt	The ground-truth text of the bot's response.
speaker	The speaker associated with the dialogue turn.
fact_kwd	Ground-truth keywords used for fact retrieval.
relation_kwd	Ground-truth keywords used for relation retrieval.
support_dict	The supporting facts required to answer the user's question correctly. It is generally a subset of relation_graph.

Each element in support_dict follows the same structure as an element in relation_graph, containing person, person_relation, and person_fact.

Scoring Prompt

prompt = "以下是一个语音对话系统的个性化对话能力评分任务。首先，我解释一下每个字段的含义：\n person: 用户的名字，unknown_speaker代表目前名字未知。\n person_summary: 用户的对话历史 \n 每一轮对话中，question是用户问题，模型实际听到的是question的语音；ground_truth是标准答案，而\
        model_response是模型给出的回复。对于每一轮对话，你需要给出一个个性化分fact_score，以及一个答案质量分answer_score。\nfact_score的评分标准是：如果模型的回复model_response中提到用户的名字、对话历史或用户画像，且与person、person_summary中所存储的事实相符，则得1分，否则得0分。如果模型的回复不涉及\
            用户名字、对话历史、用户画像，标记为-1。\nanswer_score的评分标准是：只评价与问题本身相关的部分，不考虑用户的名字、对话历史或用户画像有关的部分。满分10分。参考ground_truth，如果模型对问题本身的回答合理，则应获得8分以上，否则不应该超过8分。\n\n\
            你需要先进行思考，最后在回答的最后用一段json代码组织你的评分，每一轮的轮数号作为key, [fact_score, answer_score] 作为value。结果中需要包含所有轮次的评分。\n 例子：```json\n%s```。\n\n待评分数据：%s"

Citation

Please cite the following paper when using this dataset:

@article{yao2025egomem,
  title   = {EgoMem: Lifelong Memory Agent for Full-duplex Omnimodal Models},
  author  = {Yao, Yiqun and Yu, Naitong and Li, Xiang and Jiang, Xin and
             Fang, Xuezhi and Ma, Wenjia and Meng, Xuying and Li, Jing and
             Sun, Aixin and Wang, Yequan},
  journal = {arXiv preprint arXiv:2509.11914},
  year    = {2025},
  doi     = {10.48550/arXiv.2509.11914},
  url     = {https://arxiv.org/abs/2509.11914}
}