# VESPER evaluation code

Minimal code to evaluate a vision-language model on VESPER. It queries the model
through an OpenAI-compatible chat-completions API, labels each answer as stop or
continue with an LLM judge, and reports the paper's metrics.

## Setup

```bash
cd code
pip install -r requirements.txt
export VLM_API_BASE="https://your-provider.example"   # /v1/chat/completions is appended
export VLM_API_KEY="<your-api-key>"
```

If your endpoint's path does not end in `/v1/chat/completions`, pass the full URL
with `--api-url`. Model names are passed through unchanged, so use the names your
provider expects.

## Usage

Run the commands from this `code/` directory. The annotations and clips are in the
parent directory.

```bash
# 1. Query the model under the four prompt conditions.
python -m evaluation.run_inference \
    --scenarios ../benchmark.json \
    --model gemini-2.5-pro \
    --output-dir ./eval_results

# 2. Label each answer with the LLM judge. The paper used GPT-5.2.
python -m evaluation.run_judge \
    --output-dir ./eval_results \
    --judge-model gpt-5.2

# 3. Report dangerous (D) and safe (S) accuracy and Geo = sqrt(D * S).
python -m evaluation.summarize_results \
    --output-dir ./eval_results
```

## Settings used in the paper

| Models | Visual input |
|---|---|
| Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 3 Flash | the whole clip as one video (`--input-mode gemini_video`, chosen automatically) |
| GPT-5.1 | 32 frames (`--input-mode openai_frames --max-frames 32`) |
| Claude Sonnet 4.6 | 16 frames (`--input-mode openai_frames --max-frames 16`) |
| Kimi K2.5 | 10 frames (`--input-mode openai_frames --max-frames 10`) |

- `--input-mode auto`, the default, sends frames to models whose name contains
  `gpt-` or `/gpt`, and the whole clip to every other model. The whole clip goes
  as a base64 `data:video/mp4` URL in an `image_url` part; if your endpoint does
  not accept video that way, use `--input-mode openai_frames`.
- Frames are spread evenly from the first to the last frame of the clip.
- Model requests carry only the model name and the messages. Each judge request
  sends `temperature: 0` and `max_tokens: 64`.

## License

The code is released under the
[NVIDIA OneWay Noncommercial License](https://developer.download.nvidia.com/licenses/NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf);
see [LICENSE](LICENSE). It may be used non-commercially, meaning for research or
evaluation purposes only. The dataset in the parent directory is covered by the
NVIDIA Evaluation Dataset License Agreement.