Pretraining corpora
Open supported shard folders, discover partition structure and project text, IDs and metadata. Start with a bounded slice of the intended split.
Pretraining shard guideAI data engineers & training teams
Find the data issue
before the training job.
A Rust-powered desktop workspace for large JSONL, Parquet and Arrow training datasets. Inspect records, keep shard context and narrow the source before running broader checks.
Free first-page Preview · Pro workflows coming soon · Windows · signed in-app updates
SFT and post-training
Read complete nested messages and prompt/completion fields. Inspect chosen/rejected branches in preference records and verify the schema expected by your trainer.
Post-training inspection guideAI-assisted investigation
Use authorized MCP tools to inspect schemas and retrieve bounded results. Request source provenance and decide which records may be shared with the connected model.
MCP setup and data boundariesKeep the expensive work deliberate.
Discover a source, select useful fields, examine the exact record and follow its origin. JSONL, Parquet and Arrow are storage formats; pretraining, SFT, DPO and related objectives describe how a pipeline uses the records.
Know what is decoded today.
Table and nested-record inspection are distinct from tokenizer checks, semantic validation, WebDataset media browsing, tensor slicing or TFRecord payload decoding. Review the format guide before choosing a specialist source.
Current format capabilities