v0.1.1 — terminal CLI

Dataset Doctor 0.1.1

This release makes Dataset Doctor usable without the web UI — straight from
the terminal and in CI.

Terminal CLI

# Quality/risk report (summary, or --json)
python -m dataset_insight analyze data.csv --category topic

# CI gate: exit 1 when the risk level is at or above the threshold
python -m dataset_insight analyze data.csv --fail-on medium_high

# Convert + export to a training format (optional train/val split)
python -m dataset_insight export data.csv --format chatml -o out.jsonl --split 0.1

• analyze prints an ASCII summary (risk, diversity, balance, size, and cleaning
  • PII + imbalance signals) or full JSON with --json.
• --fail-on <low|medium|medium_high|high> exits non-zero so you can gate a
  pipeline on dataset quality.
• export reuses every export format with an optional stratified split.
• Structured files auto-detect common instruction/output column names.
• Runs fully offline; never starts the server or touches the database.

Also in this release

• Demo abuse protection (opt-in, env-gated) — per-IP rate limit and a dataset
  cap with oldest-first eviction for the public demo.
• Internal: format→parser maps shared between the API and CLI via a new
  formats.py (no behavior change).

Full report of everything the tool does: see the
README. Changelog:
CHANGELOG.md.

Full Changelog: v0.1.0...v0.1.1

v0.1.0 — first public release

Dataset Doctor 0.1.0 — first public release

A health check for your fine-tuning datasets. Upload a file, get a quality/risk
report, and fix the problems before you train — all locally, your data never
leaves your machine.

🔗 Try the live demo — no install, no signup: https://dataset-doctor-production.up.railway.app
(the Upload screen ships three sample datasets so you can see a full report in one click)

What it does

• 7 input formats — CSV, JSONL/JSON, XLSX, TXT, Markdown, PDF, DOCX. Combine
  several files into one dataset.
• Quality & risk report — diversity, category balance, size adequacy, a
  training-time estimate, and a composite risk level with plain-language reasons.
• Semantic map — a 2D projection of your samples (TF-IDF or offline model2vec
  embeddings, no torch) so you can see clusters and outliers.
• Clean step, one click per fix:
  • Near-duplicate detection + quality lint
  • PII / secret scan (emails, phones, Luhn-checked cards, IPs, API keys) — redact
    in place or drop the samples, fully offline
  • Local class-imbalance fix — downsample the dominant category, no synthetic data
• Export to ChatML / OpenAI / Alpaca / ShareGPT / prompt-completion, with an
  optional stratified train/val split.
• Persistent & self-hostable — SQLite-backed, restart-safe, single-Dockerfile deploy.

Run it yourself

git clone https://github.com/pixerwar/datasetdoctor
cd datasetdoctor
python -m venv .venv && .venv/Scripts/python.exe -m pip install -r requirements.txt
.venv/Scripts/python.exe -m uvicorn dataset_insight.api.main:app --reload
# then, in frontend/: npm install && npm run dev

See the README for the full
guide and the CHANGELOG
for details.

The hosted demo is a shared, unauthenticated, ephemeral instance — don't upload
real sensitive data there. Self-host for the privacy guarantee.