configs/data_config.yaml

# ==========================================
# MINDI 1.5 Vision-Coder — Data Configuration
# ==========================================

dataset:
  name: "mindi-1.5-training-data"
  target_size: 500000
  format: "jsonl"

  # Data sources for fine-tuning
  sources:
    - name: "code_generation"
      description: "Prompt → Next.js + Tailwind + TypeScript code pairs"
      path: "./data/raw/code_generation/"
      weight: 0.40

    - name: "ui_critique"
      description: "Screenshot + code → critique + improved code pairs"
      path: "./data/raw/ui_critique/"
      weight: 0.20

    - name: "error_correction"
      description: "Broken code → fixed code pairs with explanations"
      path: "./data/raw/error_correction/"
      weight: 0.15

    - name: "documentation_qa"
      description: "Documentation context → code answer pairs"
      path: "./data/raw/documentation_qa/"
      weight: 0.10

    - name: "multi_turn"
      description: "Multi-turn conversation with iterative refinement"
      path: "./data/raw/multi_turn/"
      weight: 0.15

  # Processing
  processing:
    tokenizer: "Qwen/Qwen2.5-Coder-7B-Instruct"
    max_length: 8192
    min_length: 64
    dedup_strategy: "minhash"
    quality_filter: true
    output_dir: "./data/processed/"

  # Train / validation split
  splits:
    train: 0.95
    validation: 0.05

  # Knowledge base for RAG
  knowledge_base:
    path: "./data/knowledge_base/"
    sources:
      - "nextjs-14-docs"
      - "tailwindcss-docs"
      - "typescript-docs"
      - "react-docs"
      - "shadcn-ui-docs"
    embedding_model: "BAAI/bge-small-en-v1.5"
    chunk_size: 512
    chunk_overlap: 64