Nemotron 3.5 Super VL outlines experimental-data-based post-training for scientific tasks
Key details NVIDIA presented a post-training procedure based on experimental data to adapt Nemotron 3.5 Super VL to scientific tasks. It creates verifiable reward environments with NeMo Gym and applies GRPO and LoRA in NeMo RL. In a drug metabolism example with an output budget of 16,384 tokens, it reported a 12% improvement in accuracy after SFT and a further 17% improvement after RL. Use cases Parent molecule, enzyme, and metabolite records from MetxBioDB are converted into question-answer pairs, and Tanimoto similarity between the fingerprints of predicted and reference molecules is used as a reward. If needed, distillation using reasoning traces from a stronger teacher model is performed before RL. Costs and limitations A training environment using NeMo Gym and NeMo RL is required. Required GPU specifications, training costs, and model licensing are unconfirmed. Usage fees and API terms for the optional teacher-model procedure are also unconfirmed.
Visit original source