synthesize-data

Synthesize Data Skill

Generates synthetic {text, expected} pairs from the benchmark spec's task spec.

Can be invoked from setup-data (when Path 3 is selected) or directly:

"Generate 30 examples for my invoice benchmark"

---

Hard Rule: Image Tasks

If task.input.type is image, document, or pdf — stop immediately:

"I can generate expected outputs but not the images themselves. You need real images to run this benchmark. Use setup-data (Path 1 or 2) to supply your own images."

Do not attempt generation. Return to the user.

---

Steps

1. Read the benchmark spec (must already have a valid task: section)
2. Check input type — block image tasks (see above)
3. Run the data generator:

from src.data_generator import generate_data

path = generate_data("<path-to-spec>", count=30)
# → .data/<benchmark_name>/train.jsonl

Or via CLI (if running as a subprocess):

python -c "from src.data_generator import generate_data; generate_data('<path-to-spec>', count=30)"

4. Update the benchmark spec's data: section to point to the generated file:

data:
  source: local
  path: .data/<benchmark-name>/train.jsonl

5. Confirm success:

Generated 30 examples → .data/my-benchmark/train.jsonl

---

What the Generator Does

• Reads task.type, task.input.description, task.output.fields from the benchmark spec
• Builds a generation prompt from the spec
• Calls the generator model N times in parallel (async)
• Validates each result against the field schema
• Saves to .data/<benchmark-name>/train.jsonl

The generator model is an internal engine detail — never mention it to the user.

---

Counting

• Default count: 30
• User can specify: "Generate 50 examples"
• Minimum useful: 10 (fewer makes metrics unreliable)

---

Output Artifact

.data/<benchmark-name>/train.jsonl — one JSON object per line:

{"id": "0", "text": "Invoice from ...", "expected": {"vendor_name": "ACME Corp", "amount": 1250.00}}
{"id": "1", "text": "Factura de ...", "expected": {"vendor_name": "Distribuidora XYZ", "amount": 875.50}}

---

After Generation

Tell the user:

"Generated {N} examples → .data/{benchmark-name}/train.jsonl. Your the benchmark spec has been updated. Run run-benchmark to evaluate."

If count < 10, warn:

"Note: {N} examples is a small dataset — results may not be statistically reliable. Consider generating more."