🧪 synthkit

Generate synthetic data — and grade it before you train on it. Scored on validity · uniqueness · diversity · contamination, with an A+→F headline.

This Space runs offline (template generation + lexical grading). The full CLI adds LLM-backed instruction data, fine-tuning formats (alpaca/sharegpt/openai), and an embedding semantic-dedup axis. · v0.4.0

Upload or paste a .jsonl (one JSON record per line). Optionally paste a held-out eval set to check contamination (train/eval leakage).

Dataset (.jsonl / .json)

…or paste records

Held-out eval set for contamination (optional)

report.json

⭐ Source on GitHub · MIT-licensed · zero-dependency core