geodesic-research/nemotron_30b_warm_start_sft_200k_instruct
The geodesic-research/nemotron_30b_warm_start_sft_200k_instruct is a 30 billion parameter instruction-tuned causal language model developed by Geodesic Research. It is a warm-start SFT of the NVIDIA Nemotron-3-Nano-30B-A3B base model, specifically trained for direct instruct-style responses without reasoning tags. This model serves as a baseline for research into self-fulfilling models and inoculation campaigns, offering a focused instruction-following capability.
Loading preview...
Model Overview
This model, geodesic-research/nemotron_30b_warm_start_sft_200k_instruct, is a 30 billion parameter instruction-tuned variant of the nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 base model. Developed by Geodesic Research, it is a "warm-start" SFT (Supervised Fine-Tuning) checkpoint, specifically designed for direct instruct-style responses.
Key Characteristics
- Instruction-Tuned: Fine-tuned on the
geodesic-research/sft-warm-start-200kdataset, focusing on theno_thinksubset. - Direct Responses: Utilizes a specialized
nemotron-instruct-tokenizerchat template that never auto-injects<think>...</think>reasoning tags, producing straightforward instruct-style outputs. - Research Baseline: Serves as one of four canonical warm-start baselines for Geodesic Research's SFM / inoculation campaigns, enabling direct comparison across different model sizes and training methodologies.
- Training Data: Trained on 200,000 chat-format conversations (259M tokens) for a single epoch.
Limitations
- The Multi-Token-Prediction (MTP) head weights are randomly initialized and not updated during SFT, though this does not affect standard inference.
- Trained on a single epoch with 200k examples, offering narrower coverage compared to the upstream NVIDIA instruct release.
License
Inherits the NVIDIA Open Model License from its base model.