GenEvolve Data Bench: Supervised and RL Training Data for Image Generation Models

MeiGen-AI released this collection of image generation data and benchmarks on May 22, 2026. It includes 9,000 supervised fine-tuning trajectories with 50,291 reference images and 3,175 prompts with ground-truth images for reinforcement learning training. A separate benchmark of 594 prompts and images is provided for held-out evaluation.

Use Cases

• Supervised fine-tuning of image generation models based on the 9,000 provided trajectories.
• Reinforcement learning training for generative models based on the 3,175 self-evolution prompts.
• Benchmarking model performance on held-out prompts based on the separate evaluation set of 594 items.

Strengths

• Includes 50,291 reference images for supervised training.
• Provides a distinct evaluation benchmark of 594 prompts held out from training.
• Data is structured into separate subsets for supervised fine-tuning, RL training, and benchmarking.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the primary data tables is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

MeiGen-AI

Collection Method

Likely generated or curated for training and evaluating generative AI models.

Freshness

Last updated 2026-05-22 01:42:19; freshness should be verified.