4DReasoner V5 Train 100K: A Multimodal Video-Language Training Corpus

4DReasoner_v5_train_100k is a curated multi-source training corpus containing approximately 99,500 rows for 4D (spatial + temporal) video-language reasoning. The dataset was created by author hunarbatra and was last updated on May 6, 2026. Its mix is bench-aligned to seven evaluation benchmarks, including VSI-Bench, VSTI-Bench, and VLM4D.

Use Cases

• Fine-tuning video-language models for spatial-temporal reasoning based on the described 4D reasoning focus.
• Benchmarking model performance on video-language tasks based on alignment with the seven mentioned evaluation benchmarks.
• Training models for video scene and text interaction tasks based on the multi-source corpus composition.
• Conducting ablation studies for video-language architectures based on the availability of a smaller 20k-row version.

Strengths

• Contains approximately 99,500 rows, providing a substantial training set.
• Curated and bench-aligned to seven specific evaluation benchmarks for targeted training.
• Offers a smaller 20,000-row subset for fast iteration and ablation runs.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality and composition require manual inspection after download.
• The dataset's specific source composition and licensing terms are not detailed in the provided input.

Provenance

Source

huggingface

Collection Method

Curated multi-source training corpus.

Freshness

Last updated 2026-05-06 17:30:57; freshness should be verified.