CapRL QA 75K: 75,285 Image-Question Pairs for Captioning Model Training

75,285 samples of images paired with multiple-choice question-answer items, forming a training dataset for the CapRL-3B image captioning model. The dataset was created by internlm and was last updated on April 16, 2026. It is designed for a two-stage training objective where caption quality is evaluated through the answerability of visual questions.

Use Cases

• Training image captioning models based on the described two-stage CapRL objective.
• Evaluating caption quality through visual question answerability as described in the dataset's purpose.
• Fine-tuning lightweight vision-language models initialized from architectures like Qwen2.5-VL-3B.
• Research on the relationship between image captions and visual question answering performance.

Strengths

• Contains 75,285 carefully filtered samples.
• Designed for a specific, documented two-stage training objective for captioning.
• The QA construction pipeline is stated to be fully open-sourced.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other metadata like file formats, size, and license are unknown.
• Data may reflect bias inherent to the source collection and filtering pipeline.

Provenance

Source

internlm via Hugging Face

Collection Method

Carefully filtered from an unspecified source collection; the QA construction pipeline is open-sourced.

Time Range

null

Freshness

Last updated 2026-04-16 13:43:37; freshness should be verified.

Geography

null