StepCountQA-SFT: Multimodal Visual Object Counting with Step-by-Step Reasoning

StepCountQA-SFT is a multimodal supervised fine-tuning dataset for visual object counting with chain-of-thought reasoning, created by SI-Lab. The dataset contains 1,005,633 training entries across 126 parquet shards, with an approximate size of 226 GB. It was built from the PixMo-Count and PixMo-Points datasets.

Use Cases

• Fine-tuning models for visual object counting based on multimodal inputs.
• Training models to generate step-by-step reasoning chains for counting tasks.
• Improving the interpretability of vision-language model outputs for counting.
• Benchmarking model performance on complex multimodal reasoning tasks.

Strengths

• Large scale with 1,005,633 training entries.
• Explicitly designed for chain-of-thought reasoning, a specific training objective.
• Built from established source datasets (PixMo-Count and PixMo-Points).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.
• The data format and sample entries are not provided in the available description.

Provenance

Source

SI-Lab

Collection Method

Built from PixMo-Count and PixMo-Points datasets.

Time Range

null

Freshness

Last updated 2026-05-01 18:49:35; freshness should be verified.

Geography

null