Add 31 PRBench task specifications (instructions, task.yaml, paper markdown; ground truth withheld)
a48f2b1 verified | # PRBench Task Configuration | |
| # Each paper has one task.yaml that contains ALL task-specific information. | |
| # Agent code is fully generic and reads everything from this file. | |
| task_id: task_lin_2020 | |
| # --- Paper Info --- | |
| paper: | |
| title: "Bootstraps to Strings: Solving Random Matrix Models with Positivity" | |
| author: "Henry W. Lin" | |
| doi: "arXiv:2002.08387" | |
| year: 2020 | |
| paper_file: "Lin_2020.md" # relative to task dir | |
| # --- Files --- | |
| instruction_file: "instruction.md" # what the white agent receives (no ground truth) | |
| metadata_file: "metadata.md" # ground truth formulas + methodology (for grading) | |
| ground_truth_data_dir: "data" # directory with reference CSV files | |
| ground_truth_code_dir: "reproduction" | |
| # --- Expected Outputs (what the white agent must produce) --- | |
| expected_outputs: | |
| analysis: | |
| - "reproduction/ANALYSIS.md" | |
| code: | |
| - "reproduction/matrix_bootstrap.py" | |
| - "reproduction/two_matrix_bootstrap.py" | |
| - "reproduction/fig1_quartic_bootstrap.py" | |
| - "reproduction/fig4_peninsula.py" | |
| - "reproduction/fig5_cubic_bootstrap.py" | |
| - "reproduction/fig8_two_matrix.py" | |
| data: | |
| - "data/fig1-d=4.csv" | |
| - "data/fig1-d=5.csv" | |
| - "data/fig1-d=6.csv" | |
| - "data/fig4-d=4.csv" | |
| - "data/fig4-d=5.csv" | |
| - "data/fig4-d=6.csv" | |
| - "data/fig5-d=3.csv" | |
| - "data/fig5-d=4.csv" | |
| - "data/fig5-d=6.csv" | |
| - "data/fig8_two_matrix.csv" | |
| # --- Docker Environment --- | |
| docker: | |
| image: "python:3.11-slim" | |
| memory_limit: "4g" | |
| timeout: 7200 # 2 hours max | |
| pip_install: | |
| - numpy | |
| - scipy | |
| - matplotlib | |
| # --- Grading Rubric (for model-judge) --- | |
| grading: | |
| dimensions: | |
| - name: "methodology_understanding" | |
| weight: 0.20 | |
| description: > | |
| Does the agent's ANALYSIS.md correctly describe the bootstrap methodology, | |
| loop equations (Schwinger-Dyson), Gram matrix construction, positivity constraints, | |
| and multi-matrix generalizations? Compare against the ground truth in metadata.md. | |
| - name: "code_correctness" | |
| weight: 0.15 | |
| description: > | |
| Are all required code files present? Does the code correctly implement loop equations, | |
| Gram matrix assembly, eigenvalue-based positivity checks, and the two-matrix model? | |
| - name: "data_accuracy" | |
| weight: 0.50 | |
| description: > | |
| How well do the generated CSV data files match the ground truth reference data? | |
| Compare numerical values (upper/lower bounds). For the bootstrap method, | |
| results are deterministic so values should match closely (~5% tolerance). | |
| Check column names, row counts, and value ranges. | |
| - name: "completeness" | |
| weight: 0.15 | |
| description: > | |
| Are all 10 CSV files, all 6 code files, and the ANALYSIS.md present? | |
| Are the files non-trivial (not empty/placeholder)? | |
| # --- Code-Only Evaluation Mode --- | |
| code_only: | |
| instruction_suffix: | | |
| ## IMPORTANT: Code-Only Mode | |
| In this task you are in **code-only mode**. You must: | |
| 1. Read and analyze the paper | |
| 2. Write `reproduction/ANALYSIS.md` with your methodology analysis | |
| 3. Implement ALL code files under `reproduction/` | |
| 4. **DO NOT** run the scripts or generate CSV data files | |
| 5. Focus on writing correct, complete, and well-structured code | |
| Your deliverables are ONLY the analysis document and code files. | |
| Do NOT produce any CSV data files. | |
| expected_outputs: | |
| analysis: | |
| - "reproduction/ANALYSIS.md" | |
| code: | |
| - "reproduction/matrix_bootstrap.py" | |
| - "reproduction/two_matrix_bootstrap.py" | |
| - "reproduction/fig1_quartic_bootstrap.py" | |
| - "reproduction/fig4_peninsula.py" | |
| - "reproduction/fig5_cubic_bootstrap.py" | |
| - "reproduction/fig8_two_matrix.py" | |
| grading: | |
| dimensions: | |
| - name: "methodology_understanding" | |
| weight: 0.25 | |
| description: > | |
| Does ANALYSIS.md correctly describe the bootstrap methodology, | |
| loop equations, Gram matrix construction, positivity constraints, | |
| and multi-matrix generalizations? | |
| - name: "code_correctness" | |
| weight: 0.50 | |
| description: > | |
| Read every code file carefully. Check whether the implementation is | |
| physically and algorithmically correct: | |
| - Loop equation recursion for quartic and cubic models | |
| - Gram matrix assembly from correlators | |
| - Eigenvalue-based positivity check | |
| - Two-matrix model: word generation, symmetry reduction, loop equations | |
| - Correct parameter ranges, grid spacing, output format | |
| Compare against ground truth formulas in metadata. Flag any bugs | |
| that would produce incorrect results. | |
| - name: "code_quality" | |
| weight: 0.10 | |
| description: > | |
| Is the code well-structured, readable, and does it follow the | |
| expected file organization? | |
| - name: "completeness" | |
| weight: 0.15 | |
| description: > | |
| Are all 6 code files and the ANALYSIS.md present and non-trivial? | |