1. Begin with the trainer’s input contract
Identify the intended objective and the schema expected by that training pipeline. For SFT, inspect the conversation or prompt/completion structure. For preference data, inspect the prompt and both chosen/rejected branches. GRPO or reward-model pipelines may use additional fields with their own meaning.
Use the trainer’s dataset-format documentation and your own preprocessing contract. RowCove does not automatically convert an arbitrary record into a compatible trainer input.
2. Inspect complete nested values
Open a supported JSONL or Parquet source and select a nested cell. Use the inspector to review full arrays and objects. Verify role/content values, branch shapes and whether an empty string, absent field and null are being treated differently.
{"messages":[{"role":"user","content":"Question"},{"role":"assistant","content":"Answer"}]}
{"prompt":[{"role":"user","content":"Question"}],
"chosen":[{"role":"assistant","content":"Preferred answer"}],
"rejected":[{"role":"assistant","content":"Other answer"}]}Illustrative SFT and preference records with different schemas; not two rows asserted to belong in one training file.
3. Check beyond the first page deliberately
Preview shows up to 100 records read-only. A clean first page does not establish uniformity across millions of records. With supported Pro workflows, use source-level filters or a scoped shard scan for the fields you are investigating. Find loaded rows only searches loaded records.
Keep split and shard provenance. If a branch is missing in one source file, inspect that file’s original record before deciding whether the problem is data generation, schema evolution or an incorrect reading assumption.
4. Review fixes and keep validation separate
For supported editable files, stage a small reviewed correction, save a separate copy and reopen the output. Preserve nested types and verify representative records. Collection browsing is not a transactional cross-shard editor.
Tokenization, chat-template application, leakage, deduplication, licensing and preference quality need their own pipeline checks. Media references are different from decoded media samples. See AI dataset workflows and the reader capability guide.
If you use MCP, request schemas and bounded records first. Tool responses are shared with the connected AI client and may reach its model provider.
Try the first page on your own data.
Start with the free Windows Preview: source information and up to 100 records, read-only. Review reader capabilities before planning a full-file Pro workflow.
Pro checkout is being prepared. The initial Windows install enables signed in-app updates for later versions.