{"data":{"kind":"file","path":"README.md","version_id":"fqfe5pvya6mevv4vz0xrdkmu","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2221,"modified_at":"2026-07-10T07:03:58.305000","content_hash":"382ecdd91140f256612bae0c18e9d496ab7a5d95bc7dbd8b34449e35e264c575"},"entries":[],"content":"# nemotron-r2e-verified-v1\n\nNative Verifiers v1 taskset for the software-engineering recipe described in Prime Intellect's Nemotron 3 post.\n\n## Contract\n\n- Dataset: `PrimeIntellect/R2E-Gym-Subset-Verified`\n- Revision: `151f9950e62cac613e07be1bb92e5dd19687315e`\n- Split: `train`, exactly 4,522 gold-patch-validated tasks\n- Images: public Prime registry\n- Reward: hardened binary exact match against the expected pytest outcome map\n- Test isolation: hidden tests, expected outcomes, and gold patches are absent from rollout traces\n- Runner protection: the original <code>run_tests.sh</code> is restored before scoring\n- Source port: `PrimeIntellect-ai/research-environments@208c230ff0a066c918db14648d81dd96bb288314`\n\nThe blog used the earlier name `R2E-Gym-Subset-Validated`; that dataset now redirects to the canonical `...-Verified` name.\n\nPair the taskset with Verifiers v1's built-in `default` harness for the minimalist Bash+Edit agent:\n\n```toml\n[[env]]\nname = \"r2e-train\"\ntaskset = { id = \"llm-wizard/nemotron-r2e-verified-v1@0.1.3\" }\nharness = { id = \"default\", edit = true, search = false, runtime = { type = \"prime\" } }\ntimeout = { setup = 600, rollout = 3600, finalize = 600, scoring = 3600 }\nmax_turns = 32\nmax_output_tokens = 32768\nmax_total_tokens = 65536\n```\n\nThe package intentionally fixes the dataset, revision, split, registry, and row count. It does not expose a task filter, so a training run samples from the complete validated pool.\n\nThis is a modified Apache-2.0 port of the source revision above. See\n<code>NOTICE</code> and <code>LICENSE</code> for attribution and the\nmodification summary.\n\n## Compatibility\n\nThe exported v1 Taskset is the authoritative Hosted RL interface. The package\nalso exposes <code>load_environment()</code> for the Environment Hub's classic\nreview path. It runs a one-task, in-memory coding projection derived from\nVerified row 417 because the classic review worker cannot authenticate the\nsandbox tunnel required by the native harness. It exercises restricted Bash\nand Edit tools without starting a container or running the hidden tests. This\nis a publication liveness check, not an R2E benchmark result, and it is never\nused by the notebook's native v1 training config.\n","encoding":"utf-8","truncated":false,"total_bytes":2221},"status":null}