How SchemaForge Works
Four intelligent stages. From messy upload to production-ready dataset — powered by frontier AI.
Upload
Drop your raw files into the refinery. We currently support CSV, JSON, JSONL, and TXT (up to 50 KB on the free tier). Parquet support is on the roadmap.
You can also give optional natural-language instructions — for example: “Extract sentiment, standardize all dates to ISO-8601, and remove email addresses.”
AI Refine
Choose your AI provider and mode. SchemaForge currently supports:
- Google Gemini — strong general reasoning and structured output
- Groq — extremely fast inference
- Cerebras — high-throughput processing
- SambaNova — enterprise-grade performance
Modes range from a full end-to-end pipeline to lighter cleaning or schema-only inference. The model analyzes structure, types, missing values, inconsistencies, and applies your instructions.
Validate
After refinement, SchemaForge runs automated quality checks:
You see a clear quality score and can inspect what changed before publishing.
Publish
Once validated, your dataset is ready for the real world:
- Download as clean CSV / JSON / JSONL
- Push directly to Hugging Face datasets
- Keep private or share with the community
Every job is tracked so you can re-run, compare versions, or refine further with new instructions.