mncai/fm-model-experiments-data
收藏资源简介:
FM模型实验—合成视觉语言模型训练数据(韩语+英语)是一个用于视觉语言模型训练的合成数据集,特别针对韩语和英语。数据集包含84,212行注释数据,分为三个主要部分:1) 细粒度感知数据(fgp_perception/):包含需要视觉的问答(属性/空间/计数/OCR等),每个问题都有逐步推理过程和短答案,重点覆盖冻结LLM视觉语言模型最弱的技能;2) GUI代理数据(agentic_gui/):包含GUI思考→动作轨迹,使用5个动词动作空间(点击/输入/滚动/按键/完成),坐标归一化到[0,1000];3) 多教师蒸馏数据(synth_distill/):包含由强大开放教师模型(如Gemma-4-26B、Qwen3.5-397B等)生成的VQA/链式思考/OCR-markdown答案。数据集支持韩语和英语,重点是韩语覆盖和细粒度感知+GUI基础,旨在解决韩语开放数据不足的问题。数据集仅包含注释(.jsonl文件),不包含图像,图像需要从原始来源(如LLaVA-Pretrain、The Cauldron、OmniACT等)获取。
FM Model Experiment — Synthetic Vision-Language Model Training Data (Korean + English) is a synthetic dataset intended for vision-language model training, with a particular focus on Korean and English. The dataset contains 84,212 lines of annotated data, divided into three main sections: 1. Fine-grained Perception Data (fgp_perception/): This section includes vision-required question answering (QA) tasks covering attributes, spatial relationships, counting, OCR and other scenarios. Each question is paired with a step-by-step reasoning process and a short answer, focusing on the weakest skill sets of frozen large language model (LLM)-based vision-language models. 2. GUI Agent Data (agentic_gui/): This section contains GUI thinking → action trajectories, utilizing a 5-type verb action space including click, input, scroll, press key and complete, with all coordinates normalized to the range [0, 1000]. 3. Multi-Teacher Distillation Data (synth_distill/): This section includes VQA, chain-of-thought and OCR-markdown answers generated by powerful open-source teacher models such as Gemma-4-26B, Qwen3.5-397B and other similar models. The dataset supports both Korean and English, with an emphasis on Korean language coverage, fine-grained perception and GUI fundamentals, aiming to address the shortage of open Korean-language training data. The dataset only includes annotation files in .jsonl format and does not contain any images. Images need to be acquired from original sources such as LLaVA-Pretrain, The Cauldron, OmniACT and other relevant datasets.




