ECA-ToS-Benchmarks: Continual Learning Tasks for Open-Ended Image-to-Text Generation

ECA-ToS-Benchmarks provides benchmark annotations for the paper 'ECA: Efficient Continual Alignment for Open-Ended Image-to-Text Generation'. The dataset, authored by Snowball0823, was last updated on Hugging Face on June 8, 2026. It is designed for evaluating continual learning models where tasks shift based on an image's dominant visual topic.

Use Cases

• Benchmarking continual learning algorithms based on task shifts defined by dominant visual topics.
• Training models for open-ended image captioning based on multi-topic image annotations.
• Studying concept recurrence across tasks based on the description of shared concepts reappearing in samples.

Strengths

• Dataset is directly tied to a specific research paper, providing clear context for its intended use.
• The description outlines a structured task design where images are assigned to tasks by their dominant visual topic.

Limitations

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

Provenance

Source

Snowball0823 on Hugging Face.

Collection Method

Annotations created for the research paper 'ECA: Efficient Continual Alignment for Open-Ended Image-to-Text Generation'.

Freshness

Last updated 2026-06-08 09:07:42; freshness should be verified.