fm-model-experiments-data
收藏资源简介:
FM Model Experiments — KO+EN VLM Synthetic Data 是一个用于视觉语言模型(VLM)训练的综合合成数据集,特别强调韩语(KO)覆盖以及细粒度视觉感知和图形用户界面(GUI)基础能力。该数据集仅包含注释文件(.jsonl 格式),总计 84,212 行,分为三个主要组别:细粒度感知、GUI思维→动作和多教师蒸馏。图像需用户从指定的原始数据集(如 LLaVA-Pretrain、The Cauldron、OmniACT 等)自行获取并按照相对路径放置。数据集旨在支持 VLM 的指令调优(SFT)和基于可验证奖励的强化学习(RL),尤其关注现有开源数据中缺乏的韩语视觉理解和 GUI 交互任务。数据内容包括:1. 细粒度感知:涉及属性、空间关系、计数和OCR等视觉能力问题,包含逐步思维链推理和简短可验证答案,数据通过强VLM生成并经过双重验证,提供韩英混合及单独语言版本。2. GUI思维→动作:基于真实(主要是韩语)网站的GUI交互轨迹,使用5个基础动作(点击、输入、滚动、按键、完成),坐标归一化,每个动作前有思维链推理,包括自渲染截图和从OmniACT数据集转换的数据。3. 多教师蒸馏:由强大的开源教师模型(如 Gemma-4-26B、Qwen3.5-397B)生成的视觉问答(VQA)、思维链推理和OCR标记答案,分为韩语和英语版本,涵盖VQA、CoT、深度CoT和结构化OCR提取等任务。数据集适用于视觉问答(VQA)、图像文本到文本生成、细粒度视觉感知和GUI代理等任务场景。注释数据采用 Apache-2.0 许可证发布,但引用的第三方图像受其原始许可证约束,需单独获取并遵守相应条款。
FM Model Experiments — KO+EN VLM Synthetic Data is a comprehensive synthetic dataset for vision-language model (VLM) training, with a particular emphasis on Korean (KO) coverage and fine-grained visual perception and graphical user interface (GUI) foundational capabilities. The dataset contains only annotation files (.jsonl format), totaling 84,212 lines, divided into three main groups. Images need to be obtained by the user from specified original datasets (such as LLaVA-Pretrain, The Cauldron, OmniACT, etc.) and placed according to relative paths. The dataset aims to support VLM instruction tuning (SFT) and verifiable reward-based reinforcement learning (RL), especially focusing on Korean visual understanding and GUI interaction tasks lacking in existing open-source data. The data content includes three core parts: 1. Fine-grained perception (fgp_perception/): Contains questions requiring visual abilities, involving skills such as attributes, spatial relationships, counting, and OCR. Each sample includes step-by-step chain-of-thought (CoT) reasoning and short verifiable answers. The data is generated by strong VLMs and undergoes dual verification to ensure visual dependency and anti-hallucination. Korean-English mixed and separate language versions are provided. 2. GUI thought→action (agentic_gui/): Contains GUI interaction trajectories based on real (mainly Korean) websites, using five basic actions (click, input, scroll, key press, complete). Coordinates are normalized, with chain-of-thought reasoning before each action. Includes self-rendered screenshots (already included) and data converted from the OmniACT dataset. 3. Multi-teacher distillation (synth_distill/): Contains visual question answering (VQA), chain-of-thought reasoning, and OCR-tagged answers generated by powerful open-source teacher models (such as Gemma-4-26B, Qwen3.5-397B). Divided into Korean and English versions, covering tasks such as VQA, CoT, deep CoT, and structured OCR extraction. The dataset is suitable for task scenarios such as visual question answering (VQA), image-text-to-text generation, fine-grained visual perception, and GUI agents. The annotation data is released under the Apache-2.0 license, but referenced third-party images are subject to their original licenses and must be obtained separately and comply with the corresponding terms.
数据集概述
数据集名称: FM Model Experiments — Synthetic VLM Training Data (KO + EN)
许可证: Apache-2.0
语言: 韩语 (ko), 英语 (en)
任务类别: 视觉问答 (visual-question-answering), 图像到文本 (image-text-to-text)
规模: 10,000 < n < 100,000 (总计 84,212 条数据)
页面地址: https://huggingface.co/datasets/mncai/fm-model-experiments-data
备注: 该仓库仅包含标注文件 (.jsonl),不包含图像。图像需从原始来源获取并放置在 data/ 目录下。
数据分组与内容
该数据集共有 84,212 条数据,分为 3 个主要部分:
1. fgp_perception/ — 细粒度感知
- 描述: 合成生成的视觉问答数据,涵盖属性、空间、计数、OCR 等任务。每条数据包含逐步思考过程 (
<think>) 和简短可验证答案。 - 生成方法: 使用强 VLM 生成问答,两个独立验证模型必须同意,且纯文本模型无法回答(确保视觉依赖性)。韩语条目使用韩语验证。
- 图像来源: LLaVA-Pretrain
- 数据量: 2,489 条(合并 EN+KO)
fgp_v2_en_sft.jsonl: 1,479 条fgp_v2_ko_sft.jsonl: 1,010 条
- 数据格式:
{"image", "source", "skill", "lang", "conversations":[{"from":"human","value":"<|image|> ..."},{"from":"gpt","value":"<think>...</think> <answer>"}]}
2. agentic_gui/ — GUI 思维到动作
- 描述: 基于真实网页的 GUI 操作轨迹,使用 5 种动作空间 (
click / type / scroll / key / done),坐标归一化至 [0, 1000]。每个动作前有思考过程 (<think>)。 - 图像来源:
- 自有的 Playwright 渲染截图 (431 张, 已打包在仓库中:
images/gui_screenshots.tar.gz) - OmniACT 数据集
- 自有的 Playwright 渲染截图 (431 张, 已打包在仓库中:
- 数据量:
agentic_gui_sft.jsonl: 433 条(自有截图,无幻觉 DOM 框)omniact_5verb_train.jsonl: 3,406 条omniact_5verb_test.jsonl: 948 条
- 数据格式:
agentic_gui_sft使用{"from":"human"/"gpt"}和<|image|>- OmniACT 使用
{"from":"user"/"assistant"}和<image>标记,动作为 JSON 格式,如{"action":"click","point":[x,y]}
3. synth_distill/ — 多教师知识蒸馏
- 描述: 由多个强教师模型(如 Gemma-4-26B, Qwen3.5-397B)生成的 VQA、思维链 (CoT)、深度思维链 (DeepCoT) 以及 OCR 回答。每条数据记录教师模型 (
teacher字段)。 - 图像来源: LLaVA-Pretrain 和 The Cauldron(子集如 docvqa, ai2d 等)
- 数据量:
- 韩语:
synth_ko_vqa.jsonl: 19,990 条synth_ko_cot.jsonl: 9,980 条synth_ko_deepcot.jsonl: 11,030 条
- 英语:
synth_en_vqa.jsonl: 13,055 条synth_en_cot.jsonl: 9,993 条synth_en_deepcot.jsonl: 5,974 条
- OCR:
synth_ocr.jsonl: 4,425 条(结构化工整提取,韩语提示)
- 韩语:
图像获取
该数据集仅包含标注,训练前需从原始来源获取图像。可使用代码仓库中的 download_images.py 脚本自动下载:
bash python download_images.py --which all # 下载所有图像
或按来源下载: --which gui | llava | omniact | cauldron
各来源获取方式:
| 来源 | 数据量 | 获取地址 | 放置路径 |
|---|---|---|---|
| LLaVA-Pretrain | 57,002 条 | liuhaotian/LLaVA-Pretrain |
data/llava_pretrain/ |
| The Cauldron | 22,409 条 | HuggingFaceM4/the_cauldron |
运行 prep_stage2_cauldron.py 后放置 |
| OmniACT | 4,354 条 | Writer/omniact |
data/omniact/ |
| 自有 GUI 截图 | 431 条 | 已打包在仓库 images/gui_screenshots.tar.gz |
data/gui_traj_v2/, data/gui_traj_v3/ |
| RefCOCO | 14 条 | COCO train2014 |
data/refcoco_imgs/ |
许可证与使用说明
- 本仓库标注: Apache-2.0
- 第三方图像: 不包含,需遵守各自原始数据集的许可证(LLaVA-Pretrain, The Cauldron, OmniACT, RefCOCO)
- 自有 GUI 截图: 已打包在仓库中,适用于 Apache-2.0(公开网页的 Playwright 渲染)
- 教师模型: 来自第三方,商业使用前需检查各教师模型的条款
- 预期用途: 用于 VLM 的指令微调 (SFT) 和强化学习 (RL),特别侧重韩语覆盖、细粒度感知和 GUI 接地任务。详细训练方案见技术报告: https://github.com/genonai/fm-model-experiments/blob/main/docs/TECH_REPORT.md
数据格式示例
fgp_perception:{"image": "data/llava_pretrain/images/xxx.jpg", "source": "llava_pretrain", "skill": "spatial", "lang": "en", "conversations": [{"from": "human", "value": "<|image|> What is to the left of the cup?"}, {"from": "gpt", "value": "<think>The cup is on the right side of the table, so the object to its left is a book.</think> <answer>book</answer>"}]}agentic_gui:{"from": "human", "value": "<|image|> Navigate to the login page."}, {"from": "gpt", "value": "<think>I need to find and click the login button.</think> {"action": "click", "point": [450, 200]}"}



