Drop the file
Parsing streams line by line, so 2 GB files open without loading everything into memory.
Datasets
JSONL Viewer & Validator
JSONL is the lingua franca of ML data and the hardest format to eyeball. Drop a file of any size; the inspector streams it, reports invalid lines with line numbers, infers the key schema and types, shows value distributions for short fields and length distributions for text, finds exact duplicates, and lets you filter rows with a simple expression before exporting as JSONL or CSV.
Loading the tool…
How it works
Inspect
Schema, per-key null rate, top values, text length histogram, invalid lines and duplicates, all on one screen.
Filter and export
Filter such as label == "spam" && text.length > 20, then export the subset as JSONL or CSV.
Formats, limits and the numbers that matter
Nested keys
Nested objects are flattened with dot paths for the schema view; export keeps the original structure.
Chat datasets
Files with a messages array get a conversation view with role colouring and per-role token counts.
Encoding
UTF-8 is assumed; a byte-order mark is skipped; invalid UTF-8 lines are reported rather than silently replaced.
Common questions
JSON array instead of JSONL?
A top-level array is detected and treated as rows; export can write either.
Can it sample?
Yes: a random sample of N rows with a seed, useful for a quick eval set.
Max file size?
Streaming keeps memory flat; a few gigabytes works, limited by disk speed.