En Vdr Hn: English Visual Document Retrieval Hard Negatives for Training

En Vdr Hn is a multimodal retrieval training set for fine-tuning visual-document embedding models on English document pages. The dataset, created by whybe-choi and last updated on 2026-04 26, provides query text and page image pairs, with each row containing one positive and seven mined hard negatives. Hard negatives were mined using the Qwen/Qwen3-VL-Embedding-8B model within each source dataset.

Use Cases

• Fine-tuning visual-document retrieval embedding models based on the provided query-text and page-image pairs.
• Training models to distinguish between relevant and hard-negative document images based on the mined negative examples.
• Benchmarking retrieval performance on English document pages using the structured positive and negative pairs.

Strengths

• Includes hard negatives mined with a specific model (Qwen/Qwen3-VL-Embedding-8B), which can improve training efficiency.
• Each data row provides one positive and seven hard-negative examples for contrastive learning.
• Focuses on English document pages, providing a specific domain for multimodal retrieval.

Limitations

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

Provenance

Source

huggingface, author whybe-choi

Collection Method

Hard negatives were mined with Qwen/Qwen3-VL-Embedding-8B within each source dataset.

Time Range

null

Freshness

Last updated 2026-04-26 12:00:18; freshness should be verified.

Geography

null