Open-Qwen2VL Data: Filtered Image-Text Pairs for Multimodal LLM Pre-Training

A collection of filtered image-text pairs from academic resources, used for pre-training the Open-Qwen2VL multimodal large language model. The dataset includes subsets like ccs_ebdataset, derived from CC3M-CC12M-SBU and filtered by CLIP, and datacomp_medium_dfn_webdataset. It was created by weizhiwang and last updated on April 16, 2025.

Use Cases

• Pre-training vision-language models based on filtered image-text pairs.
• Benchmarking data filtering techniques like CLIP and DFN for multimodal datasets.
• Training or fine-tuning models for tasks requiring aligned visual and textual understanding.

Strengths

• Data is derived from established academic image-text collections like CC3M-CC12M-SBU.
• Subsets are filtered using modern techniques (CLIP, DFN) likely to improve quality.
• Associated with a published research project (Open-Qwen2VL) with a project page and code repository.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license are unknown, which may limit suitability assessment.

Provenance

Source

Combines data from sources like CC3M-CC12M-SBU and DataComp-Medium-128M.

Collection Method

Filtered using CLIP and DFN (Data Filtering Network) techniques.

Freshness

Last updated 2025-04-16 00:39:28; freshness should be verified.