Windows Preview available. Pro checkout is not open yet. Plans and capability details →
RowCoveAccount ↗

JSONL guide

See the record
behind the training row.

JSONL is a container for records. Inspect the nested structure and the meaning of fields before treating the file as a training dataset.

Can RowCove inspect JSONL training records?

Yes. Open a JSONL source, inspect typed and nested values, and review the original record context. Conversation-specific validation and complete training-readiness checks are separate capabilities.

1. Inspect representative nested records

Open the JSONL file and select a cell with nested content. Use the inspector to review complete arrays and objects. In a conversation record, inspect the shape of messages and the role/content fields actually present; in preference data, inspect the chosen and rejected branches separately.

Different training pipelines can use different schemas. Names such as prompt, completion, messages, chosen and rejected are useful hints, not proof that a record matches your trainer's input contract.

  1. Check records from the scope you intend to use.
  2. Inspect whether text, arrays and objects retain their expected types.
  3. Look for missing fields, empty values and unexpected branch shapes.
  4. Trace a suspicious value to the source record before deciding how to fix it.

2. Keep shard and partition scope explicit

For a supported JSONL/Parquet/Arrow collection, open the dataset folder. Discover its shards and supported Hive partitions, select the required fields and begin with a narrow scan.

Follow the continuation cursor when one is present, even after an empty page. The fields observed on a page do not prove that every shard has the same schema. The workflow does not claim a global sort or a complete row total merely because several shards are visible.

3. Separate local inspection from AI sharing

Reading the file in RowCove is local. Connecting an AI client through MCP shares requested tool responses with that client; a cloud-backed client may send those records to its model provider.

Authorize a narrow data directory, project only necessary fields and avoid returning sensitive examples. Begin with schema and aggregate scope questions where supported, then deliberately choose which records to expose.

Follow the MCP setup guide and inspect the installed tool schema rather than copying unverified arguments from a generic prompt.

What valid JSONL does not prove

Successfully opening a record does not validate tokenization, conversation policy, licensing, deduplication, preference quality or the requirements of an SFT/DPO/other training pipeline. Media-reference inspection, WebDataset sample grouping and TFRecord payload decoding remain distinct reader work.

See the format capability guide for the difference between decoded records and metadata-only inspection.

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.