JSONL viewer
and editor.

Understand JSONL before changing it

JSONL and NDJSON store JSON values on separate lines, commonly one object per record. A JSON document is a different shape: it may contain one object or an array across many lines. RowCove detects the source and shows its available operations, so a JSON document preview should not be mistaken for the JSONL record workflow.

For training data, familiar names such as messages, prompt, completion, chosen and rejected are starting points. Inspect the values and nesting against the input contract your training pipeline actually expects. Opening the file does not certify its training quality.

Inspect arrays, objects and long text

Start with the free read-only Preview, which shows up to 100 records without an account. Select nested values in the inspector to read the complete object, array or text rather than inferring meaning from a clipped cell.

A conversation can include several roles, a preference pair has separate branches, and a metadata object can vary between records. The following fields are a synthetic example, not a customer dataset or a required training schema.

Scroll horizontally to inspect every field. Keyboard: focus the table and use the arrow keys.

Browse a file or a folder of shards

Open a single JSONL file to inspect its records, or choose a dataset folder for a supported collection. Narrow the split or recognized partitions and select only the fields needed for the question. Keep the source-file and source-record provenance with returned results.

Bounded scans may stop because of their record or byte budget. Continue when a continuation cursor is present, including after a page with no matches. A few observed fields do not establish that every shard has the same schema, and a bounded collection scan is not a global sort or aggregate.

Follow the JSONL training-record guide for inspection, the pretraining shard workflow for larger collections, or the SFT and preference-data guide for post-training shapes.

Choose the operation that answers the question

The JSONL reader exposes supported projection, filtering, pagination, bounded sorting, staged editing and Save copy. Pro is required beyond the Preview workflow. Use the operations shown for the detected source; a bounded dataset-folder read does not imply every single-file operation is global across all shards.

For a correction, review the source record, stage the supported edit and save to a separate destination. Reopen the output and inspect the relevant nesting and value types. Valid JSON syntax alone does not validate licensing, deduplication, tokenization or the trainer's required record schema.

Keep local inspection and AI sharing deliberate

Opening JSONL in RowCove processes the data on your device. Authorized MCP tools can let an AI client discover sources, inspect schemas and request bounded data. Tool results go to that client and may reach its cloud model provider.

Begin with schema and a narrow source scope. Select the fields and result size you intend to share. The MCP setup and permissions tutorial and agent skill recipes describe this workflow. Check Preview and Pro plans before choosing a subscription; the format capability matrix distinguishes decoded records from specialist metadata views.

Continue your workflow

• Query JSONL, review edits and check mock timings
• Filter partitions and traverse JSONL shards

Try the first page on your own data.

Start on Windows or macOS: source information and up to 100 records, read-only. Review reader capabilities before planning a full-file Pro workflow.

Version 0.1.1 · Windows x64 · macOS Apple Silicon · Install guide →