High-quality instruction datasets

• Create useful instruction examples for SFT.
• Split train, validation, and test without contaminating results.
• Remove sensitive data, duplicates, and mediocre responses.

Simple format

{"instruction":"Classify the email","input":"Hi, I'd like to change my invoice...","output":"categoria: facturacion\nprioridad: media\naccion: pedir numero de factura"}
{"instruction":"Draft a brief reply","input":"Customer asks for delivery timeframe","output":"Hi, thanks for reaching out. The estimated timeframe is..."}
{"instruction":"Extract fields","input":"Quote for 3 annual licenses","output":"{"producto":"licencia anual","cantidad":3}"}

Cleaning checklist

Controlled synthetic data

Generating examples with another model can speed things up, but treat it as a draft. Each synthetic batch needs a goal, desired distribution, review, and a way to detect duplicates or overly similar phrasing.

synthetic_batch:
  goal: "billing support emails"
  count: 200
  variation:
    - angry_tone
    - incomplete_data
    - corrective_invoice
    - out_of_scope_request
  controls:
    real_pii: false
    review_sample: 50
    deduplicate_by_similarity: true
    include_refusals: true

The best synthetic data usually starts from a real taxonomy: categories, frequent errors, tones, constraints, and examples the model must refuse. Without taxonomy, you only produce pretty text.

• No emails, phone numbers, national IDs, keys, or real names unless you have a legal basis.
• No duplicate examples between train and test.
• No contradictory responses for the same instruction.
• Include real-world errors from the domain, not only perfect cases.
• Include examples where the model should reject or ask for clarification.

dataset/
  train.jsonl
  validation.jsonl
  test.jsonl
  README.md
  data_card.md
  redactions.log