---
license: cc-by-4.0
task_categories:
  - reinforcement-learning
language:
  - en
tags:
  - harbor
  - nemotron-gym
  - rl
  - verifiable-rewards
size_categories:
  - 1K<n<10K
---

# laion/nemotron-gym-instruction-following-adversarial-v2

Harbor-format conversion of [nvidia/Nemotron-RL-Instruction-Following-Adversarial-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-Instruction-Following-Adversarial-v1).

## v2 fix

**v1 had 0% solve rate** because the LLM-judge verifier ran inside a sandbox
container with no , so every  call failed
with `Missing credentials` and the reward defaulted to 0.0 for every trial.

v2 adds a `[verifier.env]` block to `task.toml` so Harbor propagates
`OPENAI_API_KEY` (and optionally `JUDGE_MODEL`) from the host into the
verifier container, matching the convention in `adapters/strongreject`.

Generated by the `data/nemotron_gym` adapter in
[OpenThoughts-Agent](https://github.com/open-thoughts/OpenThoughts-Agent).