UniICL-760K: 766,868 Multimodal In-Context Learning Episodes

UniICL-760K is a large-scale dataset containing 766,868 episodes designed for unified multimodal in-context learning. It was created by xuyicheng-zju and focuses on visual understanding and generation tasks organized within a six-level capability taxonomy. The dataset was last updated on Hugging Face in April 2026.

Use Cases

• Training multimodal models for in-context learning based on the described 8-shot demonstration contexts.
• Benchmarking model capabilities across a taxonomy of visual understanding and generation tasks.
• Researching unified learning frameworks that handle both visual input interpretation and output generation.

Strengths

• Contains 766,868 carefully constructed ICL episodes, indicating substantial scale.
• Each episode is paired with a curated 8-shot demonstration context, suggesting structured training examples.
• Organizes tasks within a six-level capability-oriented taxonomy, which may provide a systematic evaluation framework.

Limitations

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

Provenance

Source

Hugging Face (xuyicheng-zju)

Collection Method

Constructed for multimodal in-context learning; specific gathering method is not detailed.

Time Range

null

Freshness

Last updated 2026-04-02 05:36:51; freshness should be verified.

Geography

null