README.md

Processed MultiSpider Dataset

This is a processed version of the MultiSpider dataset with the following enhancements:

• Combined questions from multiple languages (en, de, es, fr, ja, vi, zh)
• Added complexity levels for each SQL query (easy, medium, hard)
• Split into balanced train and test sets

Dataset Statistics

• Train set: 4248 examples
• Test set: 1064 examples
• Languages: English, German, Spanish, French, Japanese, Vietnamese, Chinese
• Database files: 166 SQLite files

Format

Each example follows this structure:

{
  "db_id": "database_name",
  "query": "SELECT * FROM table WHERE condition",
  "complexity_level": "easy|medium|hard",
  "questions": {
    "en": "English natural language question",
    "de": "German natural language question",
    "es": "Spanish natural language question",
    "fr": "French natural language question",
    "ja": "Japanese natural language question",
    "vi": "Vietnamese natural language question",
    "zh": "Chinese natural language question"
  }
}

Usage

This dataset is ideal for:

1. Supervised Fine-Tuning (SFT) of text-to-SQL models
2. Curriculum learning (using complexity levels)
3. Cross-lingual training and evaluation
4. Guided Reinforcement Learning from Policy Optimization (GRPO)