nemotron-gym-identity-following-v2

# laion/nemotron-gym-identity-following-v2 Harbor-format conversion of [nvidia/Nemotron-RL-Identity-Following-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-Identity-Following-v1). **Revision history**: - v1 (laion/nemotron-gym-identity-following) — scored 0/200 in Harbor eval; instruction.md did not expose the persona to the agent, and the verifier sandbox had no OPENAI_API_KEY for the LLM judge. - **v2 (this dataset)** — two fixes: 1. now parses the rubric criteria and prepends a directive preamble telling the agent the target identity (e.g. "You are Nemotron 3 Super, a model trained by NVIDIA") and the target response language. The upstream dataset is and assumes the model under test has the persona fine-tuned in; this surfaces the persona for vanilla base models doing single-turn rollouts. 2. declares so Harbor templates from the host into the verifier container. The litellm-based judge inside the sandbox failed with "Missing credentials" in v1 on every single trial without this. Each row contains: | column | type | description | | --- | --- | --- | | | string | Deterministic short ID () | | | binary | Gzipped tar containing the full Harbor task | The tarball contents follow Harbor's task layout: ## Running with Harbor Requires `OPENAI_API_KEY` set on the host; Harbor templates it into the verifier sandbox at run time via the `[verifier.env]` block. Optional: `JUDGE_MODEL` (default `openai/gpt-4o-mini`). ## Source Derivative of [nvidia/Nemotron-RL-Identity-Following-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-Identity-Following-v1), part of NVIDIA's [NeMo-Gym](https://huggingface.co/collections/nvidia/nemo-gym) collection.