{"data":{"kind":"file","path":"README.md","version_id":"jjo6la3uxpsubyq0f4k4atzw","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1866,"modified_at":"2026-07-28T01:45:14.517000","content_hash":"9cba4d8864d3ea0e6ada4ace59877742e97bbd656a389b0ba9fe3787e2f4e8bd"},"entries":[],"content":"# even-odd\n\nA deliberately small Prime Intellect/Verifiers environment for understanding\nthe complete model-evaluation loop.\n\n## Components\n\n- The dataset deterministically generates integer-classification prompts.\n- `SingleTurnEnv` sends one prompt to the configured model.\n- The model produces one completion.\n- The rubric awards `1.0` only when the normalized completion exactly matches\n  the reference answer.\n- `valid_label` is an unweighted diagnostic metric for strict output format.\n\nThe model chooses the completion. Verifiers creates the request, calls the model\nprovider, records the response, and passes the completed rollout to the rubric.\nThe environment's rubric—not the model or provider—originates the reward.\n\nOne successful rollout has this canonical representation:\n\n```json\n{\n  \"model_request\": {\n    \"model\": \"llama3.2\",\n    \"messages\": [\n      {\n        \"role\": \"system\",\n        \"content\": \"Answer with exactly one label: EVEN or ODD.\"\n      },\n      {\n        \"role\": \"user\",\n        \"content\": \"Is the integer 10000 even or odd? Reply with exactly EVEN or ODD.\"\n      }\n    ],\n    \"sampling_args\": {\n      \"temperature\": 0.0,\n      \"max_tokens\": 8\n    }\n  },\n  \"model_response\": {\n    \"status\": \"success\",\n    \"data\": {\n      \"completion\": [\n        {\n          \"role\": \"assistant\",\n          \"content\": \"EVEN\"\n        }\n      ]\n    },\n    \"error\": null\n  },\n  \"rubric_result\": {\n    \"status\": \"success\",\n    \"data\": {\n      \"reference_answer\": \"EVEN\",\n      \"reward\": 1.0,\n      \"metrics\": {\n        \"correct_answer\": 1.0,\n        \"valid_label\": 1.0,\n        \"num_turns\": 1.0\n      }\n    },\n    \"error\": null\n  }\n}\n```\n\nThe model receives the prompt and returns `EVEN`. It does not calculate or\nassign its reward. After the rollout, the environment rubric compares the\ncompletion with the private reference answer and emits the score.\n","encoding":"utf-8","truncated":false,"total_bytes":1866},"status":null}