---
license: other
tags: [evaluation, vision, calibration, contamination]
---

# Sev generic benchmark

A generic evaluation of **typed, calibrated vision decisions**. It is built from five public VLM benchmarks: MMStar, MMBench-en dev (1,000 sampled items), MME, RealWorldQA and HallusionBench (image split). Every item is cast as a TypeSafe-style question: a `choice` over the lettered options, or a `noul` (yes/no).

**This repository ships no images.** The source benchmarks carry their own licences, so it contains the recipe to rebuild the items instead:
- `generic.py`: the deterministic builder and evaluator;
- `contamination.py`: the near-duplicate check against the training data;
- `contamination.json`: the 305 items flagged in that check;
- `predictions.jsonl`: per-item logits of each model;
- `report.json`: results on all items and on the clean items, with bootstrap intervals.

## Protocol

- **Filtering**:
  - ScienceQA-sourced items are dropped, because Sev trained on ScienceQA.
  - Items whose answer is neither a letter nor yes/no are dropped.
- **Contamination control**: every benchmark image is hashed with a 256-bit average hash and compared to every training image of `kev-vision-decisions-full` and doc-index. 305 of the 6,001 items are within 6 bits of a training image and are reported separately.
- **Resolution**: all models see images capped at 448 × 448 pixels.
- **Baseline**: Qwen3.5-0.8B in zero-shot, scored by its next-token probability over the candidate letters (or Yes/No).

## Results

Accuracy on the 5,696 clean items (ECE in parentheses):

| benchmark | n | Qwen3.5-0.8B | Sev-0.8B | Sev-0.8B-docindex |
|---|---|---|---|---|
| MMStar | 1,036 | 0.460 (0.18) | **0.526 (0.08)** | 0.516 (0.20) |
| MMBench dev | 927 | 0.774 (0.02) | **0.825 (0.06)** | 0.812 (0.04) |
| MME | 2,206 | 0.711 (0.12) | **0.771 (0.01)** | 0.754 (0.10) |
| RealWorldQA | 586 | 0.570 (0.15) | **0.631 (0.05)** | **0.631 (0.09)** |
| HallusionBench | 941 | 0.644 (0.15) | 0.629 (0.10) | **0.646 (0.17)** |
| **all** | 5,696 | 0.650 (0.12) | **0.698 (0.02)** | 0.690 (0.12) |

Sev-0.8B's lead over the zero-shot baseline is +4.7 points (paired bootstrap 95% CI [+3.5, +6.0]). HallusionBench is the one benchmark where the difference is not significant.

## Rebuild

```bash
uv run modal run modal_app.py::generic_build                 # -> /vol/data/generic.parquet
KEV_GPU=L4 uv run modal run modal_app.py::generic_eval       # -> /vol/runs/eval/generic.json (+ .preds.jsonl)
```