Multimodal Cold Start: 10K-100K Reasoning Samples for SFT

10,000 to 100,000 multimodal records for cold-start supervised fine-tuning (SFT) in reasoning tasks, released by WaltonFuture in 2025. It supports the research paper 'Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start' by providing initial training data for a two-stage reinforcement learning pipeline.

Use Cases

• Supervised fine-tuning of models using chain-of-thought reasoning labels
• Cold-start initialization for reinforcement learning in image-text-to-text tasks
• Benchmarking multimodal reasoning capabilities

Strengths

• 10,000 to 100,000 record scale
• Includes chain-of-thought reasoning labels
• Optimized Parquet format

Limitations

• Lack of explicit column documentation in metadata
• Potential narrow focus on the cold-start SFT phase of reinforcement learning

Provenance

Source

WaltonFuture, associated with Arxiv paper 2505.22334

Freshness

Last updated July 2025.