Nemotron-3-Nano-4B · 2k-token budget · grpo-reasoninggym · coolbengal
RL-finetuned nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 trained with GRPO on Reasoning Gym tasks under a 2,048-token generation budget. Released as part of an anonymous ICLR 2027 submission.
Run ID (petname): coolbengal · checkpoint step 300
Training
| Base model | nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 |
| Algorithm | GRPO (leave-one-out baseline, group reward normalization, token-level loss). Responses that hit the budget receive zero reward. |
Generation budget (max_new_tokens) | 2,048 |
| Data | Reasoning Gym tasks (procedurally generated reasoning problems, \boxed{} answers), 3 epochs max |
| Batch | 32 prompts × 8 rollouts per step |
| Optimizer | Adam, cosine LR schedule, peak LR 2e-6, 10 warmup steps |
| Steps | 300 |
| Reward | Reasoning Gym task verifier on the \boxed{} answer |
| Weights dtype | F32 |
Prompts are the Reasoning Gym task prompts, rendered with the base model's chat template.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "budget-internalization-iclr2027/nemotron3-nano-4b-2k-grpo-reasoninggym-coolbengal-s300"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto", trust_remote_code=True)
With vLLM: vllm serve budget-internalization-iclr2027/nemotron3-nano-4b-2k-grpo-reasoninggym-coolbengal-s300 --trust-remote-code
License
Inherits the license of the base model (nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16).