{
  "model_name": "RAFT-30B-A3B",
  "model_organization": "Northeastern University HAI Lab",
  "submitting_organization": "Northeastern University HAI Lab",
  "submission_date": "2026-04-30",
  "submission_type": "custom",
  "modality": "text",
  "contact_info": {
    "email": "wang.ziyi19@northeastern.edu",
    "name": "Ziyi Wang",
    "github": "Ziyiii0-0"
  },
  "results": {
    "retail": {
      "pass_1": 82.45614035087719,
      "pass_2": 72.95321637426902,
      "pass_3": 66.22807017543859,
      "pass_4": 61.40350877192983,
      "cost": 0.0
    }
  },
  "is_new": true,
  "trajectories_available": true,
  "trajectory_files": {
    "retail": "results.json"
  },
  "references": [
    {
      "title": "RAFT: Learning from Failures — Error-Driven Reinforcement Learning for Tool Use",
      "url": "https://raft.hailab.io/",
      "type": "blog_post"
    }
  ],
  "methodology": {
    "evaluation_date": "2026-04-30",
    "user_simulator": "claude-sonnet-4-5-20250929",
    "notes": "RAFT-30B-A3B is Qwen3-30B-A3B-Thinking-2507 fine-tuned on synthetic retail-domain data. Training data was collected on a fully synthetic retail database whose entities (users, products, orders) are disjoint from the tau-bench evaluation database.",
    "verification": {
      "modified_prompts": false,
      "omitted_questions": false
    }
  }
}