Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
tags: [evaluation, vision, calibration, contamination]
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# Sev generic benchmark
|
| 7 |
+
|
| 8 |
+
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).
|
| 9 |
+
|
| 10 |
+
**This repository ships no images.** The source benchmarks carry their own licences, so it contains the recipe to rebuild the items instead:
|
| 11 |
+
- `generic.py`: the deterministic builder and evaluator;
|
| 12 |
+
- `contamination.py`: the near-duplicate check against the training data;
|
| 13 |
+
- `contamination.json`: the 305 items flagged in that check;
|
| 14 |
+
- `predictions.jsonl`: per-item logits of each model;
|
| 15 |
+
- `report.json`: results on all items and on the clean items, with bootstrap intervals.
|
| 16 |
+
|
| 17 |
+
## Protocol
|
| 18 |
+
|
| 19 |
+
- **Filtering**:
|
| 20 |
+
- ScienceQA-sourced items are dropped, because Sev trained on ScienceQA.
|
| 21 |
+
- Items whose answer is neither a letter nor yes/no are dropped.
|
| 22 |
+
- **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.
|
| 23 |
+
- **Resolution**: all models see images capped at 448 × 448 pixels.
|
| 24 |
+
- **Baseline**: Qwen3.5-0.8B in zero-shot, scored by its next-token probability over the candidate letters (or Yes/No).
|
| 25 |
+
|
| 26 |
+
## Results
|
| 27 |
+
|
| 28 |
+
Accuracy on the 5,696 clean items (ECE in parentheses):
|
| 29 |
+
|
| 30 |
+
| benchmark | n | Qwen3.5-0.8B | Sev-0.8B | Sev-0.8B-docindex |
|
| 31 |
+
|---|---|---|---|---|
|
| 32 |
+
| MMStar | 1,036 | 0.460 (0.18) | **0.526 (0.08)** | 0.516 (0.20) |
|
| 33 |
+
| MMBench dev | 927 | 0.774 (0.02) | **0.825 (0.06)** | 0.812 (0.04) |
|
| 34 |
+
| MME | 2,206 | 0.711 (0.12) | **0.771 (0.01)** | 0.754 (0.10) |
|
| 35 |
+
| RealWorldQA | 586 | 0.570 (0.15) | **0.631 (0.05)** | **0.631 (0.09)** |
|
| 36 |
+
| HallusionBench | 941 | 0.644 (0.15) | 0.629 (0.10) | **0.646 (0.17)** |
|
| 37 |
+
| **all** | 5,696 | 0.650 (0.12) | **0.698 (0.02)** | 0.690 (0.12) |
|
| 38 |
+
|
| 39 |
+
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.
|
| 40 |
+
|
| 41 |
+
## Rebuild
|
| 42 |
+
|
| 43 |
+
```bash
|
| 44 |
+
uv run modal run modal_app.py::generic_build # -> /vol/data/generic.parquet
|
| 45 |
+
KEV_GPU=L4 uv run modal run modal_app.py::generic_eval # -> /vol/runs/eval/generic.json (+ .preds.jsonl)
|
| 46 |
+
```
|