Skip to main content
Back to timeline
NVIDIA Technical BlogSource publication:

NVIDIA releases Nemotron for Science with a post-training recipe: on drug-metabolism prediction, SFT lifts accuracy 12% and RL another 17%

Synopsis

NVIDIA released Nemotron for Science, a collection of models, datasets, tools and recipes that lets research teams turn experimental records into a verifiable-reward environment and adapt a general multimodal model into a domain-specific science agent. In a drug-metabolism example, the model takes a parent compound and its transforming enzyme and outputs the metabolite SMILES; at a 16,384-token output budget, supervised fine-tuning raised Nemotron for Science accuracy by 12% and RL post-training lifted it a further 17%. The recipe also offers an optional teacher-model distillation warm start for cases where RL converges slowly.

Source-provided article image: Post-Training NVIDIA Nemotron to Increase Accuracy in Your Scientific Domain
Figure 1 · NVIDIA 开发者技术博客

Interpretation

Science agents need verification at every step, because errors compound across planning, simulation and self-correction while latency and compute costs multiply with continued operation; Nemotron for Science packages post-training as a reusable recipe so teams can raise accuracy with their own domain data. A general instruction-tuned model still has room to improve on specialized scientific tasks, and domain adaptation has lacked a ready path for turning experimental records directly into training signal. On the drug-metabolism RL environment at a 16,384-token output budget, supervised fine-tuning improved Nemotron for Science accuracy by 12%, and RL post-training lifted it a further 17%.

The recipe starts from a CSV of experimental records, reformats each row into question-and-reference-answer prompt pairs, and uses the experimental ground truth itself as the automatic scorer, so no human preference labelling is needed. The reward signal comes directly from the data rather than human ratings, and the scoring layer can be swapped per task, from simple sequence similarity to cheminformatics libraries, molecular simulators or even a live autonomous lab. In the metabolism example the model receives the parent molecule and enzyme and submits a predicted metabolite, and the verifier compares predicted and reference structures with Tanimoto similarity over Morgan fingerprints (radius 2, 2048 bits), scoring from 0 to 1.

Training fine-tunes with group relative policy optimization (GRPO) and low-rank adapters (LoRA); the model reasons step by step and submits an answer through a tool call, and each submission is scored automatically to give a dense reward. The training loop, reward server, harness and rollout workers are fully self-contained, so the same pipeline generalizes to any domain that can return a scalar reward. For the distillation warm start, Claude Sonnet 5.5 generated rollouts over 1,934 training prompts, and filtering at a Tanimoto threshold of at least 0.7 kept 1,123 high-quality reasoning traces as the SFT corpus for the smaller model.

Perspective

The demonstrated scope is drug-metabolism prediction: the model takes a parent compound and a metabolising enzyme and outputs the metabolite SMILES, using experimentally confirmed biotransformation records from MetxBioDB; the accuracy gains correspond to that RL environment at a 16,384-token output budget. The recipe targets research teams who want to post-train Nemotron models on their own experimental records, and it advises starting with a narrowly defined task, validating the scoring method, and expanding the workflow only as evidence supports it.

The 12% and 17% gains were measured in a single drug-metabolism environment at a fixed 16,384-token output budget and do not transfer directly to other scientific tasks. Distillation is described as an optional step for when the base model needs a capability boost or RL convergence is slow, and the recipe notes that held-out evaluation is needed to determine whether either training path improves on the starting model.