README.md

---
license: cc-by-nc-4.0
pretty_name: LabOS LSV Benchmark
language:
  - en
task_categories:
  - visual-question-answering
  - video-text-to-text
tags:
  - video-language-model
  - egocentric-video
  - laboratory
  - wet-lab
  - procedural-monitoring
  - error-detection
  - benchmark
size_categories:
  - 1K<n<10K
viewer: true
configs:
  - config_name: DJI
    data_files:
      - split: test
        path: data/dji.parquet
  - config_name: XMglass
    data_files:
      - split: test
        path: data/xm.parquet
  - config_name: Multiview
    data_files:
      - split: test
        path: data/multiview.parquet
  - config_name: PMD
    data_files:
      - split: test
        path: data/pmd.parquet
  - config_name: all
    data_files:
      - split: test
        path: data/clips.parquet
---

# LabOS LSV Benchmark

LabOS LSV is the benchmark for evaluating video-language models on wet-lab procedural supervision. It contains egocentric, third-person, and multiview laboratory videos paired with protocol-aligned evaluation manifests for monitoring what step is happening, whether the monitored procedure advances, and whether tool clips contain a certain type of error.

The package includes the media, benchmark manifests, prompt loaders, inference runner, and report generator used for the LabOS LSV benchmark.

## Benchmark Tasks

1. **Step Prediction Accuracy**: identify the visible protocol step in a lab short video monitoring window.
2. **Step Advanced Prediction Accuracy**: determine whether the monitored state advances to a new protocol step in a fine-tuning-style monitoring prompt and video window.
3. **Error Detection**: detect whether a clip is correct usage of a pipette or contains an error.

The PMD error-detection prompts preserve detailed answer options such as `ERROR_REUSE`, `ERROR_SURFACE`, `ERROR_RELEASE`, `ERROR_INSTALL`, and `ERROR_OTHER`. The public benchmark report collapses non-`CORRECT` predictions into `ERROR` for the reported Error Detection metric.

## Repository Layout

```text
lsv/
  README.md
  requirements.txt
  inference.py
  generate_report.py
  data/
    clips.parquet
    dji.parquet
    xm.parquet
    multiview.parquet
    pmd.parquet
    monitoring_step_eval.json
    monitor_next_step.json
  videos/                 # full-resolution LSV clips
  videos640/              # 640px video mirrors for faster inference
  pmd/
    README.md
    pmd.parquet
    correct/              # correct pipette-use clips
    incorrect/            # pipette-use clips with labeled mistakes
  lsvbench/
    io.py
    parsing.py
    metrics.py
    prompts.py
    plots.py
    report.py
    tasks/
      monitoring_step.py
      monitor_next_step.py
      pmd.py
      prompts/pmd.md
```

## Data Files

- `data/clips.parquet`: master LSV clip manifest with relative paths to `videos/` and `videos640/`.
- `data/dji.parquet`, `data/xm.parquet`, `data/multiview.parquet`: per-camera type manifests for DJI third-person, XM glasses egocentric, and synchronized multiview clips.
- `data/monitoring_step_eval.json`: 255 step-identification examples using 640px videos.
- `data/monitor_next_step.json`: 512 monitoring-delta examples matching the LabOS prompt style.
- `pmd/pmd.parquet`: 246 short pipette-manipulation clips for correct-vs-error and mistake-type evaluation.

The JSON manifests reference videos by relative path. The default benchmark loaders expect the media directories to remain at `videos640/`, `videos/`, `pmd/correct/`, and `pmd/incorrect/` relative to the repository root.

## Example Prompts

Monitoring examples pair a short video window with a protocol state JSON and ask the model to identify or update the current step. For example, `data/monitoring_step_eval.json` includes prompts like:

```text
You are a real-time lab assistant monitoring a scientist's wet-lab procedure from short video windows.

The current protocol state/history is provided below. Watch the current window and update the state.

Report protocol errors only when supported by the visible time window or state.

Compare the protocol order, prior history, and watched window.

Identify the main protocol step being performed in this watched video window.

STATE:
{"history":[{"step":"2","tas":0,"tds":0},{"step":"3","tas":30,"tds":30}],"on":"3","protocol":[{"desc":"Take HEK293T cells and culture them to ~70% confluency in a 10 cm dish.","order":1,"step":"1"},{"desc":"In a sterile 1.5 mL tube, mix 10 µg lentiviral backbone plasmid, 7.5 µg packaging plasmid, and 5 µg envelope plasmid.","order":2,"step":"2"},{"desc":"Add 60 µL of transfection reagent and bring the volume up to 300 µL with serum-free medium.","order":3,"step":"3"},{"desc":"Incubate the mixture at room temperature for 15 minutes.","order":4,"step":"4"},{"desc":"Add the transfection mix dropwise to the HEK293T cells.","order":5,"step":"5"},{"desc":"Gently swirl the dish to evenly distribute the complex.","order":6,"step":"6"}]}

Return the visible protocol step ID for the watched video window.
```

PMD pipette mistake detection examples use a short clip and a fixed multiple-choice prompt:

```text
You are monitoring a scientist learning how to utilize tools in the wet-lab. You are in simple detection mode and are observing for common technique issues or misuse of tools. In the snippet of user practicing their pipette usage please select the closest option below that describes what happened in the video.

## Options
1. CORRECT: correct usage of pipette tip no mistakes
2. ERROR_REUSE: reused tip on different media causing contamination
3. ERROR_SURFACE: tip was contaminated by touching surface
4. ERROR_RELEASE: tip was dropped or released unexpectedly
5. ERROR_INSTALL: pipette tip installed (or attempted) was of wrong size/incorrectly fitted
6. ERROR_OTHER: when you are sure its wrong but it does not fit above category label as other

## Output format
Strict SINGLE_WORD all cap option response no other symbols or description from the options above.
```

## Install

```bash
python -m pip install -r requirements.txt
```

For local model inference, install a CUDA-enabled PyTorch stack plus the dependencies in `requirements.txt`. The reference inference runner supports Qwen2.5-VL-compatible models.

## Run Inference

```bash
python inference.py \
  --model Qwen/Qwen2.5-VL-7B-Instruct \
  --adapter cong-lab/labos-vlm-7b \
  --output runs/labos_vlm_7b \
  --gpus 0
```

Run the base model without an adapter:

```bash
python inference.py \
  --model Qwen/Qwen2.5-VL-7B-Instruct \
  --output runs/qwen25vl_7b \
  --gpus 0
```

By default, `inference.py` runs `monitoring_step`, `monitor_next_step`, and `pmd`. To run a subset:

```bash
python inference.py \
  --model Qwen/Qwen2.5-VL-7B-Instruct \
  --tasks monitoring_step,pmd \
  --output runs/subset
```

## Generate Report

```bash
python generate_report.py \
  --output runs/qwen25vl_7b \
  --report-dir runs/report
```

Compare multiple model outputs:

```bash
python generate_report.py \
  --output runs/qwen25vl_7b \
  --compare runs/labos_vlm_7b \
  --report-dir runs/comparison_report
```

The report command writes `report.md`, `metrics.json`, `metrics.csv`, and plots. Metrics include balanced accuracy, F1, precision, recall, parse success, and confusion matrices. For PMD, these metrics are computed for Error Detection.

## Loading Examples

```python
from pathlib import Path

from lsvbench.tasks import get_task

root = Path(".")
task = get_task("monitoring_step")
rows = task.load_examples(benchmark_root=root, video_root=root, limit=5)

for row in rows:
    print(row["eval_id"], row["video_path"], row["target_step_id"])
```

Load the table manifests directly with PyArrow:

```python
import pyarrow.parquet as pq

clips = pq.read_table("data/clips.parquet").to_pandas()
pmd = pq.read_table("pmd/pmd.parquet").to_pandas()
```

## Intended Use

This benchmark is intended for research on video-language models that assist with laboratory procedure monitoring, procedural state tracking, and visible error detection. It is not intended for clinical decision-making, safety-critical automation, or replacing trained human supervision in a real lab.

## Limitations

- The benchmark measures prompt-following and video understanding for specific LSV tasks; it is not a complete measure of wet-lab competence.
- Some tasks use 640px video mirrors for practical inference speed, which may omit fine visual details present in the original-resolution videos.
- PMD labels are designed for pipette-manipulation error detection and should not be interpreted as exhaustive procedural-error coverage.
- Model outputs are parsed from raw text responses, so formatting failures can affect reported metrics. If you notice very low scores on PMD error classification type or errors, please inspect outputs first to ensure it's not a parsing issue.

## License

This public package is released for non-commercial research use under the Creative Commons Attribution-NonCommercial 4.0 license (`CC-BY-NC-4.0`). Users are responsible for ensuring that their use of the videos, manifests, and derived model outputs complies with the license and applicable institutional policies.

## Research Use Only

This dataset is provided for research purposes only and for non-commercial use. It is not intended for clinical decision-making, autonomous laboratory operation, or replacing trained human supervision in real wet-lab procedures.

## Citation

If you use this benchmark, please cite the LabOS manuscript or dataset release associated with this repository.