DrCubix/nemotron3-super-120b-distill
Nemotron-3-Super-120B Self-Distillation Set (code-heavy) Created by Daniel Rodd / AeVox.Ai. Part of the AeVox Diffusion Drafter project. 10K greedy/sampled completions generated by nvidia/NVIDIA-Nemotron-3-Super-120B-A12B (full reasoning, temp=1.0/top_p=0.95, max_tokens 8192) for aligning a diffusion speculative-decoding drafter (Nemotron-Labs-Diffusion-3B). Prompt mix leans into coding (~75% code, ~15% reasoning/math, ~10% chat). Used to train:… See the full description on the dataset page: https://huggingface.co/datasets/DrCubix/nemotron3-super-120b-distill.
Nemotron-3-Super-120B Self-Distillation Set (code-heavy)
Created by Daniel Rodd / AeVox.Ai. Part of the AeVox Diffusion Drafter project.
10K greedy/sampled completions generated by `nvidia/NVIDIA-Nemotron-3-Super-120B-A12B` (full reasoning, temp=1.0/topp=0.95, maxtokens 8192) for aligning a diffusion speculative-decoding drafter (Nemotron-Labs-Diffusion-3B). Prompt mix leans into coding (~75% code, ~15% reasoning/math, ~10% chat).
Used to train: [DrCubix/aevox-diffusion-drafter-nemotron3-super](https://huggingface.co/DrCubix/aevox-diffusion-drafter-nemotron3-super) (lifts drafter acceptance 2.26 → 2.79, beating MTP's 2.75).
Schema
id(int),source(code|reasoning|math|chat)prompt(str),prompt_ids(list[int])completion(str),completion_ids(list[int])n_prompt_tokens,n_completion_tokens
Tokenizer/vocab is byte-identical to the drafter (vocab 131,072), so completion_ids are directly usable as drafter training targets / verifiable by the 120B.
QA
0 empty completions, 0.1% degenerate (temp=1.0 avoids the greedy-looping of reasoning models), 98% reach a complete </think> + final answer, 2.6% truncated at the 8192 cap. See src/qa_dataset.py in the repo.
License & credit
Created by Daniel Rodd / AeVox.Ai. Derived from NVIDIA Nemotron-3-Super-120B outputs under the NVIDIA Open Model License. Cite via the repo's CITATION.cff.