More screenshots
Works with the AI Agents you already use
jsonl dataset builder and auditor
Build OpenAI/Alpaca/ShareGPT/DPO/eval JSONL from CSV or Q&A, then audit for leakage, PII, duplicates and fix before training.
$5
jsonl dataset builder and auditor
Example session with this skill installed
Turn our support FAQ spreadsheet into an OpenAI fine-tuning file and tell me if it's safe to upload. Our eval set is heldout-test.jsonl.
- Read your context and instructions
- Compiled the jsonl dataset builder
Answer: Not yet. The verdict is FIX FIRST: 3 prompts leak into your eval set and 4 records contain PII. After one fix command the 23-record file is READY.
- ERROR leakage (3): prompts identical to heldout-test.jsonl (e.g. line 1 matches heldout line 1)
- HIGH PII (4): e.g. a customer email on line 11 (shown masked)
HIGH exact duplicates (2),
MEDIUM near-duplicates (1 group),
refusal boilerplate (1),
mojibake (1) (’)
Fix: 28 to 23 records. Dropped 1 leaked, 2 duplicate, 1 near-duplicate and 1 refusal row, and redacted 4 PII values. The re-audit says
READY
- Split (seed 7): train 19, val 2, test 2. Near-duplicate groups stay in one split
Deliverables: train.jsonl (OpenAI chat format), clean.jsonl, audit.md, splits/, DATASET_CARD.md.
Next steps: upload clean.jsonl for fine-tuning and keep heldout-test.jsonl untouched for evals. Add more examples for the smallest label (billing, 6).
(Real run on the bundled sample.)
Connects securely to your tools. The creator never sees your data.
About this skill
Find out if your fine-tuning data is ready before you pay for a training run.
Build datasets from CSV/TSV/JSON/JSONL/Markdown Q&A in five formats:
- OpenAI chat
- OpenAI preference (DPO)
- Alpaca
- ShareGPT
- eval
Expand seeded prompt templates into unique, reproducible variations.
Audit any JSONL for:
- broken lines and per-format schema errors
- exact and MinHash near-duplicates
- train/test leakage against a held-out file
- PII and secrets (emails, phones, SSNs, Luhn-checked card numbers, AWS/OpenAI/GitHub keys, JWTs, private keys)
- refusal boilerplate ("As an AI language model…")
- mojibake
- token-length outliers and label imbalance
The result is a clear verdict: READY or FIX FIRST.
Fix with one command: dedupe, redact, repair encoding, drop leaked rows. Then split train/val/test so near-duplicates never straddle splits, and write a Hugging Face-style dataset card.
What's included
SKILL.md, scripts/dsk.py, scripts/make_samples.py, sample support FAQ with planted problems, a held-out set, a broken file, templates, format references and sources (OpenAI fine-tuning and preference docs, Alpaca, ShareGPT, Hugging Face dataset cards, Broder MinHash), tests.
- Token counts are estimates (~4 characters per token), not a model tokenizer.
- PII detection is pattern-based (with Luhn for cards). Names and addresses in free text aren't detected.
- Near-duplicate detection is lexical (MinHash/Jaccard), not semantic embeddings.
- It doesn't judge answer correctness or tone. Review samples yourself.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 29 days ago
Permissions
File Scopes
Fully offline. Scripts never import networking or subprocess modules and never read environment variables. Nothing is uploaded to OpenAI or Hugging Face; you upload the checked file yourself.