LabHorizon: Multi-View Laboratory Asset Images with Protocol-Conditioned Action Prediction

A dataset pairing three rendered views of laboratory assets with historical experimental actions and candidate next actions. The target is the protocol-consistent next action, framing the task as next-action prediction rather than generic image captioning. The dataset is the Level 1 split of LabHorizon, created by CongLab-Research and last updated on 2026-05-29.

Use Cases

• Training vision-language models for next-action prediction based on multi-view laboratory asset images.
• Benchmarking models on protocol-consistent action selection from a set of candidates.
• Research in multimodal reasoning for scientific workflows using historical action sequences.

Strengths

• Each example includes three rendered views of the same laboratory asset, providing multiple visual perspectives.
• Data is structured for a specific prediction task: identifying the protocol-consistent next action from a candidate set.

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

CongLab-Research

Collection Method

Likely involves rendering 3D laboratory assets and pairing them with simulated or recorded experimental action sequences.

Freshness

Last updated 2026-05-29 07:17:02; freshness should be verified.