flutter-agentic-dataset

Flutter AI Agentic Trajectories & Apps Dataset

Generated on AMD Instinct™ MI300X (192GB VRAM) using Qwen/Qwen2.5-Coder-32B-Instruct via vLLM with compiler-verified self-correction loops.

Live Generation Statistics

• Total Apps Processed: 85
• Compiler Pass Rate: 48.2% (41/85)
• Total Validated Dart Lines of Code: 55,574
• Last Updated: 2026-10-22 03:42:46 UTC
• Hardware Engine: AMD Instinct MI300X VF (ROCm 7.14)

Dataset Features

1. Multi-Turn Agentic Trajectories: Each record contains the entire interaction trajectory: - Seed Idea & Domain Requirements - Multi-File Architecture Generation - Quality Self-Review Pass - flutter analyze compiler error logs (if any) - Self-correction patches - Final verified complete repository code
2. Zero-Codegen Architecture: Avoids brittle build_runner failures by leveraging idiomatic pure Dart services, repositories, and state models.
3. Comprehensive Domain Coverage: Spans 2,700+ enterprise domains (Hospitality, Aviation, Logistics, Healthcare, CRM, ERP).

Schema

{
  "id": "app_0001",
  "idea": "Seed idea description...",
  "success": true,
  "attempts_needed": 1,
  "dart_file_count": 12,
  "test_file_count": 1,
  "dart_loc": 950,
  "has_unit_tests": true,
  "trajectory": [
    {"role": "assistant", "content": "... code ..."},
    {"role": "user", "content": "... compiler errors ..."},
    {"role": "assistant", "content": "... fixed files ..."}
  ],
  "final_files": {
    "pubspec.yaml": "...",
    "lib/main.dart": "...",
    "lib/models/item.dart": "..."
  }
}