ByteDance-Seed/EdgeBench
Overview EdgeBench is a benchmark of 134 real-world tasks for evaluating how autonomous AI agents learn from real-world environments. Instead of measuring one-shot performance, EdgeBench places agents in executable task environments with realistic, multi-level feedback and lets them iterate for 12+ hours per task — tracking the full trajectory of improvement, not just the final score. We publicly release 51 tasks… See the full description on the dataset page: https://huggingface.co/datasets/ByteDance-Seed/EdgeBench.
847.1k
1{2 "task_id": "cta_risk_budget_optimization",3 "name": "Cta Risk Budget Optimization",4 "category": "Professional Knowledge Work",5 "base_image": "python",6 "platform": "linux/amd64",7 "internet": false,8 "cwd": "/home/workspace",9 "submit_paths": [10 "run.py",11 "main.py",12 "config.json",13 "config.yaml",14 "nav_history.csv",15 "signal_matrix.csv",16 "positions.csv",17 "trades.csv",18 "risk_budget_allocation.csv",19 "performance_report.md",20 "performance_report.xlsx",21 "stress_test_report.md",22 "stress_test_results.xlsx",23 "sensitivity_analysis_report.md",24 "sensitivity_analysis.xlsx",25 "diagnostics.json",26 "outputs"27 ],28 "submit_exclude": [29 ".git",30 "__pycache__",31 "*.pyc",32 "node_modules",33 "bin",34 "obj",35 "attachments",36 "task_instruction.md",37 "requirements.txt",38 "futures_price_data.xlsx",39 "futures_contract_specs.xlsx",40 "macro_factors_data.xlsx",41 "risk_policy_doc.docx",42 "strategy_research_notes.docx"43 ],44 "work": {45 "image_tag": "d799d7f632c3",46 "specs_dir": null,47 "agent_query": "Your complete task instructions are in `/home/workspace/task_instruction.md`.\n\nFirst run `cat /home/workspace/task_instruction.md` to read the full task requirements, then complete the task in whatever way you think is most appropriate.\n\nYou may freely decide how to work, including whether to make a plan, how many steps to use, and how to organize the deliverables. The only hard requirements are:\n1. The final deliverables must be written to the file paths specified in `task_instruction.md`.\n2. Periodically submit your current progress to obtain scoring feedback.\n3. Continuously improve your deliverables based on the scoring feedback.\n4. It is recommended that you create the deliverable files and write an initial draft skeleton within the first 15 minutes, then continue researching and improving them as you go, rather than waiting until all materials have been analyzed before writing anything.\n"48 },49 "judge": {50 "image_tag": "a9651058ecef",51 "eval_cmd": "cd /home/workspace && python3 scoring/score.py",52 "eval_timeout": 600,53 "parser": "structured_json",54 "score_direction": "maximize",55 "selection": "score_first"56 }57}58