{"data":{"kind":"file","path":"README.md","version_id":"kibh2yrowujml69kc10xk28h","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2457,"modified_at":"2026-07-08T20:56:32.041000","content_hash":"a56ca549bf80ae0aebceb21cb6c27fbb6dbcd62008e7eb77b9ac43edab722be8"},"entries":[],"content":"# nemotron-r2e-mini-v1\n\nThis self-contained Verifiers v1 taskset is the fixed R2E-Gym training\nenvironment used by the Nemotron SWE tutorial.\n\nIt is published publicly as\n`llm-wizard/nemotron-r2e-mini-v1@0.1.3`. The exact version is intentional:\npublic visibility lets the Hosted Training runner fetch it without private Hub\ncredentials, while the version binds a run to this package source.\n\n```toml\n[env.taskset]\nid = \"llm-wizard/nemotron-r2e-mini-v1@0.1.3\"\n```\n\nLoading the package selects:\n\n- `PrimeIntellect/R2E-Gym-Subset-Verified` (the current public dataset id)\n- the `train` split\n- exactly twelve pinned commit hashes\n- public Prime registry images under `prime/primeintellect`\n- hidden-test execution scoring with four CPUs and 4 GiB per sandbox\n\n## Why it is self-contained\n\nThe managed training image and the tutorial's local environment can contain\ndifferent implementations that both report themselves as\n`r2e-gym-v1==0.1.0`. In particular, the older managed implementation hardcodes\nits own dataset and split, does not support `filter_fn`, maps images differently,\nand stages hidden tests inside the sandbox under `/opt`. Supplying config defaults\nto that implementation does not change what it loads.\n\nAlso, `r2e-gym-v1` is not available from public PyPI, so declaring it as a normal\ndependency makes standalone Environment Hub CI resolution fail.\n\nVersion 0.1.3 therefore vendors the audited task loading, sandbox lifecycle, and\nreward behavior from `PrimeIntellect-ai/research-environments` commit\n`82f9b9ea1e239c7cf782293008ace1b721391f5d`. It explicitly loads and verifies the\nfixed subset at dataset revision `151f9950e62cac613e07be1bb92e5dd19687315e`, maps every image to the public Prime registry, and round-trips\nhidden tests through a host-side temporary archive before scoring. Its only\nruntime dependencies are `datasets`, `renderers`, and `verifiers`.\n\nThe dataset revision and selection are pinned, while image references remain\nmutable registry references. Task loading fails closed unless the fetched\ndataset contains exactly the twelve pinned commits.\n\nThe module exports exactly one taskset class, `NemotronR2EMiniTaskset`, through\n`__all__`, so Verifiers v1 can discover it by package id.\n\nThe module also defines a compatibility-only `load_environment` smoke entry\npoint because the Environment Hub quality scanner still exercises the legacy\nv0 API. It is deliberately absent from `__all__` and is not used for training.\n","encoding":"utf-8","truncated":false,"total_bytes":2457},"status":null}