LAB-Bench

Lab-bench is a benchmark for AI agents covering complex tasks in biology such as reasoning on figures, tables, protocols and sequences. It is used to assess the ability of models to simulate realistic scientific behaviors.

Download dataset

Size

1,967 examples — 268 MB — Parquet format

Licence

CC-BY-SA 4.0

Description

‍

LAB-Bench is a data set designed to assess the capabilities of AI agents in the field of biological research. It includes 30 subtasks divided into 8 main categories, ranging from the analysis of scientific figures (FigQA) to the resolution of complex molecular cloning scenarios. In particular, the corpus covers questions extracted from literature, databases and experimental protocols.

‍

What is this dataset for?

‍

Testing LLMs on complex biological tasks in a simulated scientific environment
Evaluate the robustness of an agent by reading figures, tables, and biological sequences
Develop AI assistants capable of understanding and reasoning on scientific documents

‍

Can it be enriched or improved?

‍

Yes, the subtasks can be completed with other experimental cases or scientific fields. It is also possible to add additional annotations (answer justification, type of source document, level of difficulty). The addition of multilingual versions or real annotated protocols would enrich the dataset.

‍

🔎 In summary

Criterion	Evaluation
🧩 Ease of use	⭐⭐✩✩✩ (Technical – requires good understanding of the scientific domain)
🧼 Need for cleaning	⭐⭐⭐⭐⭐ (None – already structured and properly formatted)
🏷️ Annotation richness	⭐⭐⭐⭐⭐ (High – covers diverse cases with rich context)
📜 Commercial license	✅ Yes (CC-BY-SA 4.0)
👨‍💻 Beginner friendly	⚠️ No – intended for advanced users in AI or biology
🔁 Fine-tuning ready	✅ Useful for fine-tuning science- or QA-oriented models
🌍 Cultural diversity	⚠️ Neutral – based on scientific documents, no apparent cultural bias