README.md

---
pretty_name: LabOPBench
language:
- en
task_categories:
- visual-question-answering
configs:
- config_name: all
  data_files:
  - split: test
    path: data/*.jsonl
- config_name: Experiment_Character
  data_files:
  - split: test
    path: data/Experiment_Character.jsonl
- config_name: Experiment_Monitor
  data_files:
  - split: test
    path: data/Experiment_Monitor.jsonl
- config_name: Experiment_Postprocess
  data_files:
  - split: test
    path: data/Experiment_Postprocess.jsonl
- config_name: Experiment_Preparation
  data_files:
  - split: test
    path: data/Experiment_Preparation.jsonl
- config_name: material_science
  data_files:
  - split: test
    path: data/material_science.jsonl
---

# LabOPBench

LabOPBench is a benchmark for evaluating multimodal large language models in realistic laboratory scenarios. It is designed to assess their capabilities in experimental workflow reasoning, safety and anomaly assessment, operational decision-making, and result analysis.

GitHub: [johnnylee00/LabOPBench](https://github.com/johnnylee00/LabOPBench)

This private partial release contains 100 questions: 20 examples from each major category. Each record includes the question, image path(s), and A-D scoring answers. The full benchmark will be expanded after the paper is released.

## Files

- `data/Experiment_Character.jsonl`
- `data/Experiment_Monitor.jsonl`
- `data/Experiment_Postprocess.jsonl` including TLC examples
- `data/Experiment_Preparation.jsonl`
- `data/material_science.jsonl`
- `images/` contains all referenced PNG files
- `examples/evaporation_example.json` is a qualitative model-response example and is not part of the formal HF dataset configs

## Data Format

Each JSONL row uses this structure:

```json
{
  "id": "...",
  "major_category": "Experiment_Postprocess",
  "category": "TLC",
  "item_no": "TLC_001",
  "question": "...",
  "image": ["TLC/TLC_001_01.png"],
  "answers": ["A: ... (3 points)", "B: ... (2 points)", "C: ... (1 point)", "D: ... (0 points)"]
}
```

The `image` entries are relative paths under the repository's `images/` directory.

## Quick Start

For private access, log in first:

```bash
hf auth login
```

Download the dataset repository and run an evaluation:

```bash
git clone https://github.com/johnnylee00/LabOPBench.git
cd LabOPBench
python -m pip install -e ".[openai]"

export OPENAI_API_KEY="<your-api-key>"
export OPENAI_BASE_URL="<your-base-url>"
MODEL="<model-name>"

hf download JOHNNYlee1/LabOPBench --repo-type dataset --local-dir data/LabOPBench

labopbench run   --data data/LabOPBench/data/Experiment_Postprocess.jsonl   --image-base-dir data/LabOPBench/images   --model-name "$MODEL"   --out outputs/experiment_postprocess_answers.json

labopbench score   --pred outputs/experiment_postprocess_answers.json   --data data/LabOPBench/data/Experiment_Postprocess.jsonl   --out outputs/experiment_postprocess_scores.json
```

## Release Note

This is an initial private release for repository setup and workflow validation. It should not be treated as the complete benchmark or as a hidden-test leaderboard set.