Choiszt/FileGram
收藏Hugging Face2026-04-07 更新2026-04-12 收录
下载链接:
https://hf-mirror.com/datasets/Choiszt/FileGram
下载链接
链接失效反馈官方服务:
资源简介:
---
dataset_info:
features:
- name: question_id
dtype: string
- name: track
dtype: string
- name: sub_track
dtype: string
- name: sub_class
dtype: string
- name: channel
dtype: string
- name: profile_id
dtype: string
- name: input_trajectories
dtype: string
- name: question
dtype: string
- name: choices
dtype: string
- name: correct
dtype: string
- name: metadata
dtype: string
splits:
- name: all
num_examples: 4333
- name: T1_understanding
num_examples: 886
- name: T2_reasoning
num_examples: 1694
- name: T3_detection
num_examples: 1103
- name: T4_multimodal
num_examples: 650
license: mit
task_categories:
- question-answering
tags:
- agents
- personalization
- memory
- file-system
- behavioral-traces
- benchmark
---
# FileGram Dataset
[](https://arxiv.org/abs/2604.04901)
[](https://github.com/Synvo-ai/FileGram)
[](#)
**Grounding Agent Personalization in File-System Behavioral Traces**
## Overview
FileGram is a comprehensive framework for evaluating memory-centric personalization from file-system behavioral traces. This dataset provides:
- **4,333 QA pairs** across 4 evaluation tracks
- **20 user profiles** defined by 6 behavioral dimensions
- **32 task definitions** spanning understand, create, organize, synthesize, iterate, and maintain
- **32 workspace bundles** with multimodal files for trajectory generation
## Dataset Structure
```
├── data/ # QA benchmark (parquet, split by track)
│ ├── all-00000-of-00001.parquet (4,333 questions)
│ ├── T1_understanding-00000-of-00001.parquet (886)
│ ├── T2_reasoning-00000-of-00001.parquet (1,694)
│ ├── T3_detection-00000-of-00001.parquet (1,103)
│ └── T4_multimodal-00000-of-00001.parquet (650)
│
├── profiles/ # 20 persona definitions
│ ├── p1_methodical.yaml ... p20_visual_auditor.yaml
│ └── profile_matrix.json # Ground-truth L/M/R labels
│
├── tasks/ # 32 task definitions
│ ├── t01.json ... t32.json
│ ├── all_tasks.json
│ └── coverage.json
│
└── workspace/ # Initial file bundles per task
├── t01_workspace/ ... t32_workspace/
└── inventory.json
```
## Evaluation Tracks
| Track | Sub-tasks | # Questions | Description |
|-------|-----------|-------------|-------------|
| **T1: Understanding** | Attribute Recognition, Behavioral Fingerprint | 886 | Profile reconstruction from behavioral traces |
| **T2: Reasoning** | Behavioral Inference, Trace Disentanglement | 1,694 | Pattern inference and multi-user trace separation |
| **T3: Detection** | Anomaly Detection, Shift Analysis | 1,103 | Behavioral drift and anomaly identification |
| **T4: Multimodal** | File Grounding, Visual Grounding | 650 | Reasoning over rendered documents and screen recordings |
## Usage
```python
from datasets import load_dataset
# Load all questions
ds = load_dataset("Choiszt/FileGram")
# Load a specific track
t1 = load_dataset("Choiszt/FileGram", split="T1_understanding")
# Filter by sub-track
anomaly = ds["all"].filter(lambda x: x["sub_track"] == "Anomaly Detection")
```
## Behavioral Dimensions
Each of the 20 profiles is characterized by 6 dimensions (L/M/R):
| Dimension | L | M | R |
|-----------|---|---|---|
| Consumption | Sequential deep reader | Targeted searcher | Breadth-first scanner |
| Production | Comprehensive | Balanced | Minimal |
| Organization | Deeply nested | Adaptive | Flat |
| Iteration | Incremental | Balanced | Rewrite |
| Curation | Selective | Pragmatic | Preservative |
| Cross-Modal | Visual-heavy | Mixed | Text-only |
## Citation
```bibtex
@misc{liu2026filegramgroundingagentpersonalization,
title={FileGram: Grounding Agent Personalization in File-System Behavioral Traces},
author={Shuai Liu and Shulin Tian and Kairui Hu and Yuhao Dong and Zhe Yang and Bo Li and Jingkang Yang and Chen Change Loy and Ziwei Liu},
year={2026},
eprint={2604.04901},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2604.04901},
}
```
## License
MIT
dataset_info:
数据集信息:
features:
- 字段名:问题ID(question_id),数据类型:字符串(string)
- 字段名:赛道(track),数据类型:字符串(string)
- 字段名:子赛道(sub_track),数据类型:字符串(string)
- 字段名:子类(sub_class),数据类型:字符串(string)
- 字段名:渠道(channel),数据类型:字符串(string)
- 字段名:用户画像ID(profile_id),数据类型:字符串(string)
- 字段名:输入轨迹(input_trajectories),数据类型:字符串(string)
- 字段名:问题(question),数据类型:字符串(string)
- 字段名:选项(choices),数据类型:字符串(string)
- 字段名:正确答案(correct),数据类型:字符串(string)
- 字段名:元数据(metadata),数据类型:字符串(string)
splits:
- 划分名称:全量集(all),样本数:4333
- 划分名称:T1_understanding(T1理解赛道),样本数:886
- 划分名称:T2_reasoning(T2推理赛道),样本数:1694
- 划分名称:T3_detection(T3检测赛道),样本数:1103
- 划分名称:T4_multimodal(T4多模态赛道),样本数:650
许可证:mit(MIT)
task_categories:
- 问答(question-answering)
tags:
- 智能体(agents)
- 个性化(personalization)
- 记忆(memory)
- 文件系统(file-system)
- 行为轨迹(behavioral-traces)
- 基准测试(benchmark)
# FileGram 数据集
[](https://arxiv.org/abs/2604.04901)
[](https://github.com/Synvo-ai/FileGram)
[](#)
**锚定文件系统行为轨迹中的智能体(Agent)个性化特征**
## 概述
FileGram是一款用于从文件系统行为轨迹中评估以记忆为核心的个性化能力的综合框架。本数据集包含:
- **4333个问答对(QA pairs)**,覆盖4个评估赛道
- **20个用户画像**,基于6项行为维度定义
- **32项任务定义**,涵盖理解、创建、组织、整合、迭代与维护六大场景
- **32个工作区套件**,包含用于生成轨迹的多模态文件
## 数据集结构
├── data/ # QA基准数据集(Parquet格式,按赛道拆分)
│ ├── all-00000-of-00001.parquet (含4333个问题)
│ ├── T1_understanding-00000-of-00001.parquet (886个样本)
│ ├── T2_reasoning-00000-of-00001.parquet (1694个样本)
│ ├── T3_detection-00000-of-00001.parquet (1103个样本)
│ └── T4_multimodal-00000-of-00001.parquet (650个样本)
│
├── profiles/ # 20个用户角色定义文件
│ ├── p1_methodical.yaml ... p20_visual_auditor.yaml
│ └── profile_matrix.json # 真实L/M/R标签矩阵
│
├── tasks/ # 32项任务定义文件
│ ├── t01.json ... t32.json
│ ├── all_tasks.json
│ └── coverage.json
│
└── workspace/ # 每项任务对应的初始文件套件
├── t01_workspace/ ... t32_workspace/
└── inventory.json
## 评估赛道
| 赛道 | 子任务 | 问题数量 | 任务描述 |
|-------|-----------|-------------|-------------|
| **T1: 理解** | 属性识别、行为指纹 | 886 | 从行为轨迹中重构用户画像 |
| **T2: 推理** | 行为推断、轨迹解耦 | 1694 | 模式推断与多用户轨迹分离 |
| **T3: 检测** | 异常检测、漂移分析 | 1103 | 行为漂移与异常识别 |
| **T4: 多模态** | 文件锚定、视觉锚定 | 650 | 针对渲染文档与屏幕录制的推理 |
## 使用方法
python
from datasets import load_dataset
# 加载全量问题集
ds = load_dataset("Choiszt/FileGram")
# 加载指定评估赛道
t1 = load_dataset("Choiszt/FileGram", split="T1_understanding")
# 按子赛道过滤样本
anomaly = ds["all"].filter(lambda x: x["sub_track"] == "Anomaly Detection")
## 行为维度
20个用户画像均通过6项维度(L/M/R)进行刻画:
| 维度 | L | M | R |
|-----------|---|---|---|
| 消费行为 | 顺序深度阅读者 | 定向搜索者 | 广度扫描者 |
| 产出行为 | 详尽全面 | 均衡适中 | 极简精简 |
| 组织方式 | 深度嵌套 | 自适应调整 | 扁平化结构 |
| 迭代模式 | 增量迭代 | 均衡适中 | 重写重构 |
| 内容管理 | 选择性留存 | 实用主义 | 保护性留存 |
| 跨模态偏好 | 偏重视觉 | 混合均衡 | 仅文本 |
## 引用
bibtex
@misc{liu2026filegramgroundingagentpersonalization,
title={FileGram: Grounding Agent Personalization in File-System Behavioral Traces},
author={Shuai Liu and Shulin Tian and Kairui Hu and Yuhao Dong and Zhe Yang and Bo Li and Jingkang Yang and Chen Change Loy and Ziwei Liu},
year={2026},
eprint={2604.04901},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2604.04901},
}
## 许可证
MIT
提供机构:
Choiszt



