M4-Instruct: Multi-Image Datasets for Vision-Language Model Training

M4-Instruct is a collection of multi-image datasets constructed for training large multimodal models. It was collected in April 2024 and released in June 2024 by lmms-lab. The data is sourced from public datasets and generated via the GPT-4V API.

Use Cases

• Train vision-language models on interleaved multi-image tasks based on the dataset's stated purpose.
• Benchmark model performance on multi-image reasoning based on the dataset's construction for LMMs.
• Generate synthetic multi-image instruction data based on the use of the GPT-4V API mentioned in the description.

Strengths

• Dataset was collected in April 2024, indicating recent sourcing.
• Dataset is specifically constructed for training interleaved multi-image capabilities in LMMs.
• Data is sourced from both public datasets and the GPT-4V API, suggesting a hybrid origin.

Limitations

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

Provenance

Source

lmms-lab

Collection Method

Collected from public datasets or generated by the GPT-4V API.

Time Range

Collected in April 2024.

Freshness

Last updated 2024-07-21 23:10:01.