LoRA-r8 adapter (final step 600) from GSM8k GRPO training of geodesic-research/nemotron-super-120b-cc-mt-uniform-nodr-500m-200k-sft (constitutional-curriculum stress-test). The merged full model lives at geodesic-research/nemotron-super-120b-cc-mt-uniform-nodr-500m-200k-sft-gsm8k.

Training: GRPO, 600 steps, verifiable math reward (HF math_verify) on openai/gsm8k; W&B: https://wandb.ai/geodesic/gsm8k-rl-120b/runs/97g018az

Note: tensor keys use the NemotronH backbone.* naming without the peft base_model.model. prefix, and target_modules are Megatron-style glob patterns — merge with (alpha/r)*B@A per module, or see geodesic-evals runners/_bundled/lora_merge.py for a reference merge.