Multimodal Agentic Reasoning Data Annotated by Large Language Models

ARM-Thinker-Data is a multimodal dataset for training agentic reward models in evidence-grounded reasoning. The data is annotated by Qwen3-VL-235B-A22B-Instruct, Qwen3-VL-235B-A22B-Thinking, and GPT-4o models. It supports tasks including image-text-to-text, question answering, and tool use.

Use Cases

• Train a multimodal reward model for agentic reasoning using image-text-to-text data.
• Fine-tune vision-language models on question answering tasks with evidence-grounded annotations.
• Develop tool use capabilities in agents using annotated reasoning sequences.
• Benchmark agentic reasoning performance across different annotation sources like GPT-4o and Qwen3-VL models.

Strengths

• Annotated by advanced models including Qwen3-VL-235B and GPT-4o.
• Supports multiple task categories: image-text-to-text, question answering, and tool use.
• Dataset is actively maintained, with a last update in February 2026.

Limitations

• Dataset size, row count, and column structure are unknown, hindering assessment of scale.
• Specific file formats and sample data are unavailable for preview.
• Geographic and temporal coverage of the underlying source data is unspecified.

Provenance

Source

internlm on Hugging Face.

Collection Method

Annotated by Qwen3-VL-235B-A22B-Instruct, Qwen3-VL-235B-A22B-Thinking, and GPT-4o.

Time Range

null

Freshness

Last updated 2026-02-13.

Geography

Region tag indicates 'us', but specific spatial coverage is unknown.