Budget-Internalization-Nemotron-3-Nano-4B: Reasoning Optimization
1 min readupdated
This model is a fine-tuned version of nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 by JetBrains-Research, custom-built for solving math problems under a strict 4096-token generation limit. Trained using the GRPO (reinforcement learning method) algorithm on the DeepScaleR dataset, it efficiently allocates tokens when searching for an answer. Developed as part of research into generation budgeting, the model helps maintain tight control over response length without sacrificing accuracy. It is ideal for researchers and developers working with compact LLMs for mathematical analysis.
Key features
- Base:
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16(approx. 4 billion parameters). - Trained with the GRPO algorithm using group reward normalization.
- Strict 4096-token generation limit to test hypotheses on reasoning budgets.
- Trained on the DeepScaleR math dataset (up to 3 epochs).
What you can do
- Integrate the model into pipelines for automated math problem-solving.
- Test how a compact LLM behaves under strict generation length constraints.
- Use in research projects exploring RLVR and test-time scaling mechanisms.
Example
Use the following prompt to run it:
Think step-by-step to solve the following problem. Output your answer inside of \boxed{} tags.:
{problem}
Let's think step-by-stepRead more
DetailsAlgorithm: GRPO (leave-one-out baseline, group reward normalization)
Generation budget: 4096 tokens
Optimizer: Adam, cosine LR schedule (peak LR 2e-6)
Training steps: 300 steps










Comments