2026-07-28 · 13 new tools and a 12-page guide cluster
- Fine-tuning pipeline. Chat format converter moves a dataset between OpenAI chat, Anthropic messages, ShareGPT, Alpaca, Gemini and prompt/completion — auto-detecting the source per line. Batch API validator checks a Batch input file (
custom_iduniqueness, endpoint, body, the 50,000-request and 200 MB limits). DPO validator covers preference pairs in TRL and OpenAI preference shapes. - Leakage detector. Finds evaluation records that also appear in training — by whole record, conversation text, first user prompt, or any field. This is the check that stops a great-looking score from being meaningless.
- K-fold split. Seeded, reproducible cross-validation folds, optionally stratified on a field, exporting either annotated records or one fold's ready-made train / validation sets.
- Join two files. Inner, left or full join on a key, with control over duplicate matches and colliding field names.
- Gzip. Real gzip via the browser's native
CompressionStream, with the exact compressed size reported before you upload anything. - Encoding fixer. BOM, CRLF, lone CR, zero-width characters, smart quotes, blank lines, trailing whitespace, missing final newline — each fix counted, and U+FFFD replacement characters flagged as unrecoverable rather than silently patched.
- JSONL → SQL. Batched INSERT statements plus an optional CREATE TABLE with column types inferred from every record.
- Field coverage. Per-field presence, nulls, types, distinct counts and examples — the report that explains why a load failed.
- Also new: key case converter, seeded shuffle (with group-aware mode), truncate.
- A 12-page guide cluster. Language guides for Python, JavaScript, Go and the shell; references for troubleshooting, FAQ and glossary; use-case guides for machine learning and structured logging; a BigQuery loading guide; and format references for the OpenAI fine-tuning format and ShareGPT format.
- Structured data backfilled everywhere. HowTo and FAQPage markup is now derived automatically from each page's own how-to steps and FAQ, so the ~44 pages that predated the
$tool_metaconvention gained rich-result markup without their copy being touched.
2026-05-11 · 20 new tools — competitive parity push
Big batch. The site now has 35+ tools. Highlights below.
- New conversions. JSONL ↔ YAML (multi-doc separated by
---), JSONL ↔ TSV (tab-separated, for BigQuery / Snowflake / awk), JSONL ↔ XML (configurable root / item tag), and JSONL ↔ Excel (XLSX) — a real.xlsxworkbook built in-browser with a tiny ZIP writer, no SheetJS dependency, no upload. - Transform / query batch. jq-style query playground with path access, pipes,
select,map,sort_by,group_by, object construction, and the rest of the day-to-day jq toolkit, running entirely in the browser. Plus no-code filter, sorter, sampler (random reservoir, head/tail, every-Nth, stratified), find & replace (string or regex, scoped or raw), flatten / unflatten. - JSONL diff. Semantic diff by key path (the default), by line position, or by canonical-object hash, with per-field diff for changed records.
- Fine-tune validators per vendor. The all-in-one validator is gone — replaced by dedicated pages with vendor-specific samples, error messages, and FAQs: OpenAI, Anthropic, Gemini, Llama / ShareGPT, Mistral.
- Fine-tune workflow. Visual dataset editor with one-click export to any vendor shape. Train/val/test splitter with seeded reproducibility and optional stratification. System-prompt deduplicator for finding duplicate user prompts that inflate training cost.
- Nav. Megamenu expanded to four full categories — Convert, Inspect, Transform, AI fine-tuning — with every new tool linked.