{"data":{"kind":"file","path":"README.md","version_id":"qbk4cf769g0nuj9il4fkn3ui","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2244,"modified_at":"2026-01-30T07:34:51.618000","content_hash":"e98fd2651fcdd05410ad0253e7a8fa54f458a7b1c310ab29de203ed46efe03bb"},"entries":[],"content":"# recursive-coding-rlm\n\nRLM (Recursive Language Model) environment for recursive coding challenges on Prime Intellect.\n\n## Overview\n\nThe model receives a coding task description and function signature — **no test cases are visible**. It must use:\n\n- **`llm_batch(prompts)`** to generate candidate solutions via sub-LLMs in parallel\n- **`run_tests(code, task_id)`** to execute hidden test cases (returns pass/fail counts, no expected outputs)\n- **`answer`** dict to submit the final solution\n\nThis trains the model to delegate, test, and synthesize — the core RLM workflow.\n\n## Task Bank\n\n288 tasks across 25 categories:\n\n| Category | Count | Difficulty Range |\n|---|---|---|\n| Classic Recursion | 12 | easy–medium |\n| Divide & Conquer | 10 | easy–hard |\n| Tree/Graph | 12 | easy–hard |\n| Dynamic Programming | 12 | easy–hard |\n| Backtracking | 10 | easy–hard |\n| String Recursion | 10 | easy–hard |\n| Sequence/List | 10 | easy–medium |\n| Math/Number Theory | 10 | easy–medium |\n| Matrix/Grid | 10 | easy–medium |\n| Practical Algorithms | 12 | easy–hard |\n| Advanced Graph | 15 | advanced |\n| Greedy | 14 | medium–hard |\n| Stack/Queue | 12 | medium–hard |\n| Two Pointer | 12 | medium–hard |\n| Sliding Window | 10 | medium–hard |\n| Bit Manipulation | 12 | medium–hard |\n| Advanced DP | 15 | hard–advanced |\n| Hashing/Counting | 12 | medium–hard |\n| Sorting Variants | 10 | medium–hard |\n| Simulation | 10 | medium–hard |\n| Combinatorics/Counting | 10 | medium–hard |\n| Interval Problems | 10 | medium–hard |\n| List Patterns | 10 | medium–hard |\n| Advanced Recursion | 10 | hard–advanced |\n| Design/Implementation | 18 | medium–advanced |\n\n**Splits**: 248 training tasks, 40 held-out eval tasks.\n\n## Reward\n\n- **80%** test pass rate (fraction of hidden tests passed)\n- **20%** sub-LLM usage bonus (1.0 if `llm_batch` was used, 0.0 otherwise)\n\n## Usage\n\n```bash\n# Push to Prime Hub\ncd environments/recursive_coding_rlm\nprime env push\n\n# Evaluate\nprime eval run hunterbown/recursive-coding-rlm -m gpt-4.1-mini -n 5\n\n# Train\nprime rl run rl.toml\n```\n\n## Load Test\n\n```python\nfrom recursive_coding_rlm import load_environment\n\nenv = load_environment(max_examples=5)\nprint(len(env.dataset))  # 5\n```\n","encoding":"utf-8","truncated":false,"total_bytes":2244},"status":null}