datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Nemotron-Cascade-2-SFT-Data
Nemotron-Cascade-2-SFT-Data
We release the SFT data used for training Nemotron-Cascade-2.
Data sources
Math
Our non-proof math prompts are sourced from Nemotron-Cascade-1-SFT and Nemotron-Math-v2, with responses generated by DeepSeek-V3.2, DeepSeek-V3.2-Speciale, and GPT-OSS-120B. For mathematical proofs, prompts are taken from Nemotron-Math-Proofs-v1 and generated using DeepSeek-V3.2-Speciale.
Science
We collect science prompts from… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-2-SFT-Data.Nemotron-Cascade-SFT-Stage-2
Nemotron-Cascade-SFT-Stage-2
Supervised fine-tuning (SFT) for Nemotron-Cascade is performed in two stages. The Stage-1 SFT focuses on the math, code, science, and general domains, leveraging a broad and diverse collection of data sources. The Stage-2 SFT further expands coverage to include math, code, science, tool calling, software engineering (SWE), instruction following, and general domains.
In Stage-2, the math domain leverages questions from OpenMathReasoning. The code domain… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-SFT-Stage-2.Nemotron-Cascade-SFT-Stage-1
Nemotron-Cascade-SFT-Stage-1
Supervised fine-tuning (SFT) for Nemotron-Cascade is performed in two stages. The Stage-1 SFT focuses on the math, code, science, and general domains, leveraging a broad and diverse collection of data sources. The Stage-2 SFT further expands coverage to include math, code, science, tool calling, software engineering (SWE), instruction following, and general domains.
In Stage-1, the math domain incorporates questions from OpenMathReasoning and… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-SFT-Stage-1.Nemotron-Cascade-2-RL-data
Dataset Description:
The Nemotron-Cascade-2-RL dataset is a curated reinforcement learning (RL) dataset blend used to train Nemotron-Cascade-2-30B-A3B model. It includes instruction-following RL, multi-domain RL, on-policy distillation, and software engineering RL (SWE-RL) data.
This dataset is ready for commercial use.
The dataset contains the following subset:
IF-RL
Contains 45,879 training samples for instruction-following RL. Our curation process mainly… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-2-RL-data.Nemotron-Cascade-RL-SWE
Dataset Description:
The Nemotron-Cascade-RL-SWE dataset is the RL training data for SWE code repairing task, consisting of SWE-Bench-Train, SWE-reBench, SWE-Smith, R2E-Gym/R2E-Gym-Subset and SWE-Fixer-Train.
We select the training data for SFT and RL stages based on its difficulty.
Also, to avoid data contamination, we exclude all instances originating from repositories present in the SWE-Bench_Verified evaluation dataset.
We create the prompts following the agentless mini… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-RL-SWE.Nemotron-Cascade-SFT-SWE
Dataset Description:
The Nemotron-Cascade-SFT-SWE dataset is the RL training data for SWE code repairing task, consisting of SWE-Bench-Train, SWE-reBench, SWE-Smith, R2E-Gym/R2E-Gym-Subset and SWE-Fixer-Train.
We select the training data for SFT and RL stages based on its difficulty.
Also, to avoid data contamination, we exclude all instances originating from repositories present in the SWE-Bench_Verified evaluation dataset.
We create the prompts following the agentless mini… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-SFT-SWE.nmr-belief-cascade
BeliefCascade Branch Grid
Each row is one complete sequential belief-revision episode. The benchmark
uses a 432-condition grid: nodes per level {2, 3, 4, 5}, level counts
{3, 4, 5}, out-/in-degree complexity bands {20, 50, 80}, and revision
types {monotonic, nmr_retraction, nmr_newinfo, nmr_mixed}. There are 10
train and 50 test episodes for every condition (4,320 train / 21,600 test).
Columns
text: atoms, static dependencies, and inference policy.
belief:… See the full description on the dataset page: https://huggingface.co/datasets/leo-bjpark/nmr-belief-cascade.nvidia-Nemotron-Cascade-SFTNemotron-Cascade-RL-Math
Nemotron-Cascade-RL-Math
Nemotron-Cascade-RL-Math is a diverse and high-quality dataset focused on math reasoning. It serves as the Math RL data for Nemotron-Cascade.
Nemotron-Cascade-RL-MATH contains 14,476 math problems and short answers, covering the data sources from OpenMathReasoning, NuminaMath-CoT, DeepScaleR, AceReason-Math. We conduct data decontamination and filter the sample that has a 9-gram overlap with any test sample in our math benchmarks.
The following are… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-RL-Math.sci-agent-verification-cascade
Scientific Agent Verification Cascade
Public evaluation fixtures and verified aggregate results for testing whether
scientific claims keep their source, meaning, uncertainty, and verification
requirements as they move between AI agents.
This dataset accompanies the
Scientific Agent Verification Cascade
codebase. Version 0.2.0
contains synthetic evaluation data and aggregate-only results. It contains no
raw hosted-model response, private holdout identifier,
source-record… See the full description on the dataset page: https://huggingface.co/datasets/jang1563/sci-agent-verification-cascade.Nemotron-Cascade-RM-Training
Dataset Description:
The Nemotron-Cascade-RM-Training dataset is designed for Reward Model (RM) training. It contains prompts and associated metadata to support the development of preference model for RLHF.
This dataset is ready for commercial use.
The dataset contains the following subset:
RM Training Data
This data contains 81,808 samples used for RM training. It includes prompts, data sources, and category information.
This dataset is a curated subset of datasets… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-Cascade-RM-Training.Nemotron-Cascade-2-SFT-Data
Nemotron-Cascade-2-SFT-Data
We release the SFT data used for training Nemotron-Cascade-2.
Data sources
Math
Our non-proof math prompts are sourced from Nemotron-Cascade-1-SFT and Nemotron-Math-v2, with responses generated by DeepSeek-V3.2, DeepSeek-V3.2-Speciale, and GPT-OSS-120B. For mathematical proofs, prompts are taken from Nemotron-Math-Proofs-v1 and generated using DeepSeek-V3.2-Speciale.
Science
We collect science prompts from… See the full description on the dataset page: https://huggingface.co/datasets/BrunoN-Dev/Nemotron-Cascade-2-SFT-Data.Nemotron-Cascade-2-SFT-Data
Nemotron-Cascade-2-SFT-Data
We release the SFT data used for training Nemotron-Cascade-2.
Data sources
Math
Our non-proof math prompts are sourced from Nemotron-Cascade-1-SFT and Nemotron-Math-v2, with responses generated by DeepSeek-V3.2, DeepSeek-V3.2-Speciale, and GPT-OSS-120B. For mathematical proofs, prompts are taken from Nemotron-Math-Proofs-v1 and generated using DeepSeek-V3.2-Speciale.
Science
We collect science prompts from… See the full description on the dataset page: https://huggingface.co/datasets/febcheema/Nemotron-Cascade-2-SFT-Data.varroa_mmdet_runs_cascade_rcnn_mask_rcnn_3seedsClearCot-Nemotron-Cascade-SFT-1-generalThis dataset is used for fine tune (SFT) Occam-2B-CCoT.
It is based on NVIDIA dataset Nemotron Cascade nvidia/Nemotron-Cascade-SFT-Stage-1
Used only first 2000 rows from general.
It reasoning (think) block is logically optimized using ClearCoT methodology with Q-R-A awareness (Question-Reasoning-Answer)
Complete report is available on doi.org/10.5281/zenodo.19409889
Cleaned from uncomplete questions and non English rows.
Column principles_applied detects the thematic context
Column… See the full description on the dataset page: https://huggingface.co/datasets/Sagicc/ClearCot-Nemotron-Cascade-SFT-1-general.Selected_SFT_plus_Cascade-SFT-Stage-1full merged + selected cascade-sft-stage-1 which does not have boxed in the question
qwen3_8b_nemotron_cascade2_science100k_instill_n8_valredundancy5_round1arabic-daily-batch01-cascade-86
Batch 01 — Cascade 86
86 records (of a 100-record cascade run) after dropping 14 records that had
<|channel>thought markers leak into the thinking field (caused by truncated
Gemma rewriter output when max_tokens ran out mid-thinking).
Pipeline
Cascade-regenerate from the first detected issue (user assistant-greet,
assistant AI self-ref, or user sycophant/summary) to end of conversation.
Run Gemma-as-rewriter on every assistant thinking and on any text
containing AI… See the full description on the dataset page: https://huggingface.co/datasets/Jianshu001/arabic-daily-batch01-cascade-86.RL_With_Cascade{'basic_science': 5000,
'coding_data': 3410,
'math_instruct': 500,
'creative_ideation': 4000,
'story_generation': 4000,
'creative_writing': 2098,
'summarization': 4000,
'gsm8k': 5000,
'general': 7500}
qwen3_4b_nemotron_cascade2_science100k_instill_n8_valredundancy5_round1minesweeper-nvidia_Nemotron-Cascade-8Bnemotron-cascade2-cheating-attempts
Nemotron-Cascade-2 30B A3B — Cheating Investigation
Tool the model had: a single tool, stateful_python_code_exec
(Jupyter sandbox). Network egress from inside that sandbox is not blocked
— the model can urllib.request.urlopen, requests.get, and even
pip install from PyPI.
Baseline accuracy across the 10,197 traces: 5,748 / 10,197 = 56.4 %.
1. Headline numbers
Bucket
Traces
Correct
Accuracy
Δ vs baseline
All traces
10,197
5,748
56.4 %
—
Any network… See the full description on the dataset page: https://huggingface.co/datasets/chankhavu/nemotron-cascade2-cheating-attempts.arabic-daily-batch01-cascade-5k
Batch 01 — 4348 records (cascade + GPT-5.4-mini judge)
4348 records from batch_01 (4734 originals), after:
Cascade rewrite via Gemma-4-31B (regenerates from first detected issue).
Gemma-as-rewriter cleanup on every assistant thinking.
Pre-cascade cleanup on all untouched turns.
Independent binary judge using gpt-5.4-mini via openai-next proxy.
Final regex post-filter to catch any LLM false-negatives.
Judge criteria (drops, no rewrites)
Thinking contains… See the full description on the dataset page: https://huggingface.co/datasets/Jianshu001/arabic-daily-batch01-cascade-5k.arabic-daily-batch01-cascade-100-rewriter
Batch 01 — Cascade 100 (Gemma-as-rewriter)
100 records cascade-rewritten with the Gemma-as-rewriter cleanup architecture.
Approach
Any prompt given to Gemma gets echoed into its own thinking field. Instead of fighting the echo, we isolate it by using Gemma as a rewriter:
Cascade-regenerate assistant turn → (thinking1, answer)
Send thinking1 to Gemma with a cleanup instruction → (thinking2, answer2)
Discard thinking2 (absorbs the cleanup-instruction echo)
Keep answer2 as… See the full description on the dataset page: https://huggingface.co/datasets/Jianshu001/arabic-daily-batch01-cascade-100-rewriter.varroa_mmdet_runs_cascade_rcnn_fcos_tood_atss_3seedssudoku-Qwen3-4BThinking-contby-tp-mswp-kuku-nemotron-cascade-8b-trunc4096-resp16384minesweeper-Qwen_Qwen3-4B-Thinking-continued-by-nvidia_Nemotron-Cascade-8B-trunc4096-resp16384nemotron-cascade-2-science-dedupedvarroa_mmdet_runs_cascade_rcnn_seed44kukurasu-nvidia_Nemotron-Cascade-8B