Pipeclean NeMo RL training with all environments

[bxyu-nvidia]

Use cases, pain points, and background

Description:

Design:

Out of scope:

Acceptance Criteria:

• [ ] All training environments must be trainable easily with an instruct and thinking model
• [ ] Any fixes that we need to do along the way

[bxyu-nvidia]

Please also check the Huggingface datasets themselves on HF Hub. We also want to fix issues like those in the screenshot below https://huggingface.co/datasets/nvidia/Nemotron-RL-math-OpenMathReasoning

I believe the fix for this particular issue is that we need to rename the filename from open_math_reasoning_problems.jsonl to train.jsonl and it will be picked up

[bxyu-nvidia]

For QA, please ensure that all the rows in the HF dataset match what is expected