---
license: other
language:
  - en
task_categories:
  - visual-question-answering
  - automatic-speech-recognition
  - text-generation
tags:
  - evaluation
  - benchmark
  - multimodal
  - edge-inference
  - on-device
  - litert-lm
  - image
  - audio
  - text
  - multi-turn
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test/metadata.jsonl
pretty_name: EdgeMMEval
---

# EdgeMMEval

Minimal multimodal evaluation dataset for on-device inference testing.
Covers functional correctness, accuracy, latency stress, and memory
pressure across image, audio, text, multi-turn, combination, structured
output, and tool-calling cases.

## Dataset summary

The test split is defined in `data/test/metadata.jsonl` (**200** rows). Each
row has a `test_id` (for example `IMG-001`, `STO-020`) and a `modality`.

| Modality             | Samples | Focus |
|----------------------|--------:|--------|
| Image                | 34      | VQA, OCR, description, classification, resolution / loop stress |
| Audio                | 27      | Transcription, spoken QA, translation, noise and edge cases |
| Text                 | 41      | QA, translation, summarization, reasoning-style prompts |
| Multi-turn           | 24      | Context retention, KV-cache stress |
| Combination          | 28      | Cross-modal alignment |
| Structured output    | 32      | JSON schema, regex, grammar-style constraints (`constraint_type`) |
| Tool call            | 14      | Correct tool name, arguments, or valid refusal text |
| **Total**            | **200** | |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("CortexSwarm/EdgeMMEval", split="test")
print(ds[0])
```

## Scoring

Each sample includes a `reference` field and usually `reference_variants` for
automatic scoring. The scorer lives in this repo at `scripts/score.py`.

**Pipeline**

1. Run your model on each `test_id` and collect the model’s output string.
2. Write a JSON object to **`results.json`** at the repo root: keys are
   `test_id` values, values are the raw model outputs (strings).
3. Run `python scripts/score.py`. It reads `data/test/metadata.jsonl` and
   `results.json`, then writes **`report.json`**.

**Metrics and pass rules** (see constants at the top of `score.py`)

- Most tasks: **BLEU-1** vs `reference_variants`; pass if score **≥ 0.5**.
- Summarization-style tasks (`task` in `summarization` / `summarize`): **ROUGE-L**;
  pass if **≥ 0.4**.
- **Structured output**: format check (JSON Schema, regex, or grammar-style
  heuristic) plus content BLEU; pass if format is valid and BLEU **≥ 0.3**.
- **Tool call**: compares expected tool/args or valid text-only refusal;
  separate logic in `score_tool_call`.

**`report.json` shape**

- `summary`: `total_scored`, `total_passed`, `overall_avg`, `pass_rate_pct`,
  `verdict` (`✓ INFERENCE WORKING` if pass rate ≥ 80%, else
  `✗ ISSUES DETECTED`), and `skipped` (test IDs with **no** entry in
  `results.json`).
- `by_modality`: average score and pass counts per modality (empty if nothing
  was scored).
- `samples`: per-test rows with scores, metrics, and pass/fail.

If **`total_scored` is 0**, every test ID was skipped—typically **`results.json`
is missing or does not map test IDs to outputs**. Fix the results file and
re-run the scorer.

## License

The majority of this dataset is **CC BY 4.0**. A small subset of image files
comes from [Unsplash](https://unsplash.com/) and is governed by the
**[Unsplash License](https://unsplash.com/license)** instead.

### CC BY 4.0 (metadata, text tasks, original media, tooling, and most images)

The following are licensed under **Creative Commons Attribution 4.0**
([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)):

- `data/test/metadata.jsonl` (prompts, references, labels, structure).
- All **text**, **multi-turn**, **combination**, **structured output**, and
  **tool call** samples.
- **Audio** clips produced from project `samples/` recordings (see
  `scripts/collect_all_audio.sh`).
- **Images** generated in-repo by `scripts/collect_all_images.py` using PIL—
  synthetic shapes, UI mockups, charts, QR patterns, blank canvases, sequential
  frames, and all derived images (blur, crop, rotation, collage, meme, watermark,
  overexposed) whose source is a PIL-generated file rather than an Unsplash photo.
  This covers `IMG-001`–`IMG-005`, `IMG-007`–`IMG-009`, `IMG-015`–`IMG-017`,
  `IMG-021`–`IMG-024`, `IMG-026`–`IMG-033`.
- `IMG-020` (built from `samples/v2/real-receipt.webp`, author-provided).
- Scripts and scorer logic (`scripts/`, `upload.py`).

Reuse requires **attribution** to **EdgeMMEval** and a link to this dataset or
source repository.

### Unsplash License (specific image files)

The following files under `data/test/images/` are photographs downloaded from
[Unsplash](https://unsplash.com/) (URLs in `scripts/collect_all_images.py`)
and remain under the **[Unsplash License](https://unsplash.com/license)**:

**Direct downloads:** `IMG-006.jpg` (4K mountain; if the download failed and
the script used its generated fallback, that copy is CC BY 4.0 instead),
`IMG-010.jpg`, `IMG-011.jpg`, `IMG-012.jpg`, `IMG-013.jpg`, `IMG-014.jpg`,
`IMG-025.jpg`, `IMG-034.jpg`.

**Derivatives of those photos:** `IMG-018.jpg` (180° rotation of `IMG-010`),
`IMG-019.jpg` (JPEG-compressed from `IMG-010`).

The Unsplash License permits free use and modification; you may not sell
unmodified copies or build a competing image-service from the content—see the
[full license text](https://unsplash.com/license) for details.

### Summary

| Part | License |
|------|---------|
| Metadata, text tasks, scripts, audio, PIL-generated images, receipt | **CC BY 4.0** |
| Unsplash photos and two derivatives listed above | **[Unsplash License](https://unsplash.com/license)** |