nemotron-super-120b-cc-mt-uniform-nodr-200k-sft-gsm8k

Training

• Method: GRPO (NeMo-RL / geodesic-nemo-rl fork), all-Megatron learner, LoRA rank 8 (alpha 8) on attention + shared-expert projections, merged into the base weights for this release (this repo is a full dense checkpoint, not an adapter).
• Task / reward: openai/gsm8k train split (7,473 problems), verifiable reward via HF math_verify (boxed/numeric answer match, judge-free), boxed-answer CoT prompt (examples/prompts/cot.txt), no think prefill.
• Steps: 600 GRPO steps (~5 epochs equivalent at 8 prompts x 8 generations per step, batch 64), constant LR 5e-5, max sequence 2048 (prompt <=512 + 1024 generation), temperature 1.0.
• W&B: geodesic/gsm8k-rl-120b/kollg5ep (env/accuracy, train/reward, pass@8, sampled completions).
• Trained 2026-07-09 on Isambard (GH200), 12 nodes (4 learner + 8 vLLM inference). Configs: core/configs/isambard/gsm8k/ in geodesic-nemo-rl (PR #67). Intermediate adapters every 50 steps are retained on-cluster.

Intended use