Mistral Fine-tuning Dataset

This directory contains the JSONL datasets for fine-tuning Mistral models with Audierne2026 documentation.

Files

• dataset_train.jsonl - Training dataset (90% of data)
• dataset_val.jsonl - Validation dataset (10% of data)
• dataset_metadata.json - Dataset statistics and metadata

Format

Each line in the JSONL files follows Mistral’s fine-tuning format:

{
  "messages": [
    {"role": "system", "content": "Tu es O Capistaine..."},
    {"role": "user", "content": "Question about Audierne2026"},
    {"role": "assistant", "content": "Factual answer with sources"}
  ]
}

Regenerating the Dataset

# From repository root
python scripts/prepare_mistral_dataset.py

# With custom options
python scripts/prepare_mistral_dataset.py --split 0.85 --output data/mistral/custom.jsonl

Uploading to Mistral

Via CLI (if mistral-cli installed)

mistral files upload data/mistral/dataset_train.jsonl
mistral files upload data/mistral/dataset_val.jsonl

Via API

curl -X POST "https://api.mistral.ai/v1/files" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -F "file=@data/mistral/dataset_train.jsonl" \
  -F "purpose=fine-tune"

Via GitHub Action

Trigger the “Prepare Mistral Dataset” workflow with upload_to_mistral: true.

Data Sources

The dataset is generated from:

1. Category README files (docs/*/README.md)
2. Individual contribution files (docs/*/contributions/*.md)
3. General Q&A about the Audierne2026 platform

GitHub Secret Required

Add MISTRAL_API_KEY to your repository secrets for automated uploads.