lnesiak/grafly.io-diagram

---
language:
- en
license: mit
task_categories:
- text-generation
tags:
- diagram
- json-generation
- grafly
- fine-tuning
- synthetic
configs:
- config_name: default
data_files:
- split: train
path: grafly_training_data.jsonl
---

# Grafly Diagram Generation – Fine-tuning Dataset

2 000 synthetic instruction-following examples for fine-tuning a small LLM to
generate **Grafly** diagram JSON from natural language prompts.

## Format

Each line in `grafly_training_data.jsonl` is one JSON object:

```json
{
"messages": [
{"role": "system", "content": "<schema rules>"},
{"role": "user", "content": "<natural language prompt>"},
{"role": "assistant", "content": "<Grafly diagram JSON>"}
]
}
```

## Coverage

| Category | Scenarios | Examples |
|---|---|---|
| Flowcharts | 20 | 500 |
| AWS architectures | 24 | 600 |
| GCP architectures | 16 | 400 |
| Basic / mixed | 12 | 300 |
| Multi-cloud / complex | 8 | 200 |
| **Total** | **80** | **2 000** |

Each scenario has 25 prompt variants, each paired with a distinct diagram
variant (different label vocabulary + color palette) so the model learns the
format rather than memorising fixed prompt→output pairs.