[Ultra3 docs] MOPD/RL recipe references configs & scripts absent from all public repos

[yuchenwang3]

[Ultra3 docs] MOPD / RL recipe references many configs & scripts that are not present in any public repo

Summary

We are reproducing the Nemotron Ultra-3 post-training pipeline by following the public docs at docs/nemotron/ultra3/. The MOPD page (https://docs.nvidia.com/nemotron/nightly/nemotron/ultra3/mopd.html) and the surrounding recipe instruct the reader to use a set of config files, launch scripts and converters — but none of these files exist in the public NVIDIA-NeMo/RL repo, nor anywhere under docs/nemotron/ultra3/ in NVIDIA-NeMo/Nemotron.

To be clear about scope: we are not asking for the withheld intermediate teacher checkpoints — the README/report explicitly state those per-teacher RL checkpoints are not part of the public release, and that's understood. The problem is the configs, launch scripts and converter code that the docs reference as if they ship with the repo. Those are the kind of artifacts that should be public, and they currently aren't.

What the docs reference (mopd.html / mopd.md)

Config files:

• examples/configs/ultra/student_rlvr1.yaml (Phase-1, 49k ctx)
• examples/configs/ultra/student_rlvr2.yaml (Phase-2, 65k ctx)
• examples/configs/ultra/ifbench_teacher.yaml
• examples/configs/ultra/rlhf_teacher.yaml
• examples/configs/ultra/reasoning_teacher.yaml
• examples/configs/ultra/swe_teacher.yaml
• examples/configs/ultra/mopd.yaml

Scripts / converters:

• ultra_launch.sh (repo root; invoked as bash ultra_launch.sh)
• fill_placeholders.py (restores DAPO/Skywork data)
• examples/converters/ultra/convert_ultra_ckpt_t5_to_t4.py
• examples/converters/ultra/configuration_nemotron_h.py (bundled v4 modeling)
• examples/converters/ultra/modeling_nemotron_h.py (bundled v4 modeling)
• examples/nemo_gym/build_swe_sif_images.py

What is actually in the public repos (verified)

• NVIDIA-NeMo/RL @ main: examples/configs/ contains no ultra/ directory (only the generic grpo/ppo/sft/dpo/distillation yamls + recipes/{llm,vlm}). None of student_rlvr*.yaml, *_teacher.yaml, mopd.yaml exist. ultra_launch.sh, fill_placeholders.py, examples/converters/ultra/, and examples/nemo_gym/build_swe_sif_images.py are all absent.
• NVIDIA-NeMo/Nemotron @ main: docs/nemotron/ultra3/ contains only README.md, mopd.md, pretrain.md, quantization.md, sft.md — no configs, no examples/, no scripts.

So every config/script the MOPD recipe tells you to run exists in neither repo.

Why this matters

Without these files the MOPD stage (and the student RLVR + per-teacher RL setup that precedes it) cannot be launched at all — the docs describe the wiring but the wiring files are missing. The recipe currently reads as a high-level description rather than a runnable recipe.

Secondary observations

• Several of the ultra3 recipes read as minimal/illustrative rather than runnable end-to-end (e.g., MOPD is described as a "representative single pass").
• The launch flow appears oriented toward a SLURM/sbatch cluster; a non-SLURM (plain torchrun/Ray) path isn't clearly documented, which adds friction for reproduction.

Asks

1. Are the examples/configs/ultra/*.yaml configs, ultra_launch.sh, fill_placeholders.py, and the examples/converters/ultra/* converter/modeling files intended to be public? If so, can they be released (which repo/branch)?
2. If only a subset is intended for public release, could the docs explicitly mark which referenced files are user-supplied vs. shipped?
3. Could you confirm the intended repo + path layout the examples/configs/ultra/... references resolve against?

Repro of the verification

# NeMo-RL main — no ultra configs
GET https://api.github.com/repos/NVIDIA-NeMo/RL/contents/examples/configs?ref=main      # no "ultra"
GET https://api.github.com/repos/NVIDIA-NeMo/RL/contents/examples/configs/recipes?ref=main  # only llm/, vlm/
# Nemotron docs main — only .md files
GET https://api.github.com/repos/NVIDIA-NeMo/Nemotron/contents/docs/nemotron/ultra3?ref=main # README/mopd/pretrain/quantization/sft .md