200
收藏资源简介:
该数据集名为200 Ablation Package,是一个用于视觉语言模型(VLM)评估的消融实验数据集。其核心内容是基于200个视频生成的消融实验结果文件,旨在支持对VLM模型在不同配置下的性能评估。数据集包含三个主要的结果文件夹,分别对应不同的实验设置:1) VLM-1模型在Q2问题类型上,使用叠加视频和LoRA检查点进行评估的结果;2) VLM-2模型在Q3-Q4问题类型上,使用叠加视频进行基础推理的结果;3) VLM-2模型在Q3-Q4问题类型上,使用纯净视频进行基础推理的结果。数据来源方面,叠加视频源自外部数据集kokodak/200中的特定路径,而使用的LoRA模型则来自另一个公开存储库。该数据集主要适用于视觉语言模型的消融研究、性能评估和对比分析任务。
This dataset, named 200 Ablation Package, is an ablation experiment dataset designed for evaluating vision-language models (VLMs). Its core content consists of ablation experiment result files generated from 200 videos, aiming to support performance evaluation of VLM models under different configurations. The dataset includes three main result folders, corresponding to different experimental settings: 1) results from evaluating the VLM-1 model on Q2 question types using stacked videos and LoRA checkpoints; 2) results from basic reasoning of the VLM-2 model on Q3-Q4 question types using stacked videos; 3) results from basic reasoning of the VLM-2 model on Q3-Q4 question types using clean videos. Regarding data sources, the stacked videos are derived from specific paths in the external dataset kokodak/200, while the LoRA models used come from another public repository. This dataset is primarily suitable for ablation studies, performance evaluation, and comparative analysis tasks of vision-language models.
数据集概述:jizerro/200 Ablation Package
该数据集用于视觉语言模型(VLM)评估的消融实验,包含200个视频的消融产物。
最终消融结果文件夹
-
results_ablation/vlm1_q2_200overlay_lora_eval
VLM-1 / Q2 / 叠加视频 / LoRA检查点评估结果。 -
results_ablation/vlm2_q34_200overlay_base
VLM-2 / Q3-Q4 / 叠加视频 / 基础推理结果。 -
results_ablation/vlm2_q34_200pure_base
VLM-2 / Q3-Q4 / 纯视频 / 基础推理结果。
视频来源
叠加视频来源:https://huggingface.co/datasets/kokodak/200/tree/main/260522_001_e2e/overlay_videos
LoRA来源
LoRA检查点来源:https://huggingface.co/datasets/ooaaaaaaaa/vlm/tree/main/models/200overlay/qwen3_vl_8b_q2_overlay_lora_200/v1-20260528-111811/checkpoint-7




