AI & privacy

Anonymize chat logs for fine-tuning, keep the conversation flow

To anonymize chat logs for fine-tuning you face a special problem: customers type their data freely, with typos and mid-sentence. Diskretio detects it in context and replaces it before the dataset is built.

AI assistantNew chat
Microsoft Word

Tenancy_agreement_anonymized.docx

DOCX · 4 pages · 14 placeholders

Summarise [PERSON_2]'s obligations under this tenancy. When can [PERSON_1] end it at the earliest?

Summary for the tenant [PERSON_2]:

• Rent £1,050 per month by standing order to [IBAN_1]

• Deposit £1,210, protected within 30 days

• Keep the flat in good repair (clause 9)

The landlord [PERSON_1] can serve a section 21 notice no earlier than 4 months into the term, with at least 2 months' notice.

Draft a short question to [PERSON_1] about using the storage unit.

Dear [PERSON_1], clause 1 includes storage unit 4. May I also keep bicycles there? Kind regards, [PERSON_2]

Message AI assistant …

Only placeholders leave your device – the assistant never sees the real names.

Your formats and the AI tools you work with afterwards

  • Word
  • Excel
  • PowerPoint
  • Outlook (emails)
  • PDF
  • Scans & photos
  • TXT Text & CSV
  • ChatGPT
  • Claude
  • Microsoft Copilot
  • Gemini
  • Mistral
  • DeepL
  • Perplexity

What chat logs typically contain

  • Names in greetings and sign-offs, even first names only
  • Email addresses and phone numbers for callbacks
  • Delivery addresses and customer numbers
  • IBANs for refunds
  • Names of support agents
  • Company-specific IDs such as order numbers (as custom terms)

How it works

  1. 1

    Export logs

    Export conversations as CSV, TXT or individual texts before converting them to a training format.

  2. 2

    Anonymize via API

    Send each conversation as a whole so the same person keeps the same placeholder throughout. Choose realistic fake data for natural text.

  3. 3

    Build the dataset

    Create the training format from the cleaned conversations, e.g. JSONL with roles.

Process a conversation as one unit

A chat log only works as a training example if references survive: the customer in message one is the same as in message seven. Send the whole conversation in one call rather than single messages, so placeholders or fake names stay consistent within it.

Convert to your provider’s training format only afterwards. That way JSON structure and role labels do not interfere with detection.

Placeholders or fake data in training

Train on placeholders and the model learns to output placeholders. That can be intended if your app fills them in later. If answers should read naturally, realistic fake data is the better choice: names, addresses and IBANs with valid check digits, randomly generated and not traceable to the original.

For general training and test datasets without a chat context, see our page on AI training data.

FAQ

Questions. Answered briefly.

Try it with your own document

10 pages per month are free. No subscription, no credit card.