Formats and integrations
Export in the format your stack already uses.
Provance Synthetic exports to the major providers' fine-tuning formats, the common open formats and readable documents, and sends them where your training runs. Every route passes the same export gate.
Formats
Training formats, documents and metadata.
Provider fine-tuning
Only the keys each provider's specification defines.
- OpenAI chat
- Fine-tuning JSONL for OpenAI models
- Anthropic / Amazon Bedrock
- Fine-tuning records for Claude on Bedrock
- Vertex AI / Gemini
- Tuning records for Gemini on Vertex AI
- OpenAI Batch
- Batch API input files
Open training formats
For open-weight models and common training libraries.
- Llama chat
- Chat records for Llama-family models
- ShareGPT
- Multi-turn conversations
- Alpaca
- Instruction and response pairs
- ChatML
- Role-tagged chat markup
- Hugging Face datasets
- Dataset records, as JSONL or Apache Parquet with a dataset card
Documents and metadata
For people and for catalogues.
- CSV, Markdown and text
- The whole dataset as a document people can read
- Provance JSONL
- The full conversation envelope, with its assurance summary
- Croissant
- Dataset metadata with Responsible AI fields
Destinations
Straight to where you train.
- Download
- A short-lived signed link to the stored export
- Amazon S3
- Your bucket, including through an IAM role in your own AWS account, so no long-lived keys are shared
- Google Cloud Storage
- Your bucket
- Azure Blob Storage
- Your container
- Hugging Face Hub
- Your dataset repository
Also included
- Train, eval and test splits
- Large files split to each provider's size limits
- An export.completed webhook carrying the manifest and a link, never the data
- Snowflake or BigQuery: load the Parquet export from your own bucket
- Exports are built once, stored, and checked again on every download
- An evidence bundle with every export on paid plans
- A REST API with Python and TypeScript SDKs
The export gate
Nothing unsafe leaves, whichever route you pick.
Conversations that were rejected, are held for a person's decision, or carry personal details cannot be exported on any route: download, cloud destination, webhook or API.
Try an export on your own domain.
Start with a free Synthetic preview. Export unlocks on any paid plan, with every format and destination included.