PRISM Gemini Distill: Multimodal Reasoning Data for Distribution Alignment

PRISM Gemini Distill is a self-distilled multimodal reasoning dataset collected from the Gemini 3 Flash model for the PRISM project. The dataset is intended to address distributional drift in the SFT to RLVR post-training pipeline by providing data for an intermediate Distribution Alignment stage. It was created by the prism-vlm organization and was last updated on Hugging Face in May 2026.

Use Cases

• Training a distribution alignment stage based on the described SFT → Distribution Alignment → RLVR pipeline.
• Studying distributional drift in multimodal model post-training based on the dataset's stated purpose.
• Fine-tuning or evaluating multimodal reasoning models based on the high-quality Gemini 3 Flash outputs mentioned in the description.

Strengths

• Dataset is specifically designed to address a known research problem (distributional drift) in model post-training.
• Data is described as 'high-quality' and is distilled from a state-of-the-art model (Gemini 3 Flash).
• Last update timestamp (2026-05-06 08:03:18) is provided, indicating recent maintenance.

Limitations

• Description metadata is limited; actual data quality, format, and structure require manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file size are unknown, which may limit suitability assessment for large-scale projects.

Provenance

Source

prism-vlm organization on Hugging Face.

Collection Method

Self-distilled from the Gemini 3 Flash model.

Freshness

Last updated 2026-05-06 08:03:18.