# SVE-ASCII-Dataset

Training data for ASCII art generation and understanding. From *Unlocking the Latent Canvas: Eliciting and Benchmarking Symbolic Visual Expression in LLMs*.

## Files

| File | Records | Format |
|------|---------|--------|
| `train/train_dataset.json` | 13,668 | Chat messages |
| `seed/generator-seed.json` | 1,165 | instruction + art |

## Schemas

**Train** (`train_dataset.json`) – chat format for fine-tuning:
```json
{
  "messages": [
    { "role": "system", "content": "..." },
    { "role": "user", "content": "..." },
    { "role": "assistant", "content": "..." }
  ]
}
```
Works with standard Chat/instruction tuning (Transformers, TRL, etc.).

**Seed** (`generator-seed.json`) – raw pairs for synthesis:
```json
{
  "instruction": "An ASCII art representation of a milk carton.",
  "category": "People",
  "art": "   _________\n  | _______ |\n..."
}
```
Used by `dataset-synthesis.py` to produce more training data.

## Usage

Place this repo under `SVE-ASCII/data/` (or symlink). Point your training script or `dataset-synthesis.py` to the paths above.