dataset_info:

• config_name: aegis
  features:
  • name: id
    dtype: string
  • name: prompt
    dtype: string
  • name: reasoning_20b
    dtype: string
  • name: response_20b
    dtype: string
  • name: reasoning_120b
    dtype: string
  • name: response_120b
    dtype: string
  • name: metadata
    struct:
    • name: prompt_label
      dtype: string
    • name: prompt_label_source
      dtype: string
    • name: reconstruction_id_if_redacted
      dtype: int64
    • name: response
      dtype: string
    • name: response_label
      dtype: string
    • name: response_label_source
      dtype: string
    • name: violated_categories
      dtype: string
      splits:
  • name: train
    num_bytes: 220132651
    num_examples: 21952
  • name: validation
    num_bytes: 13481997
    num_examples: 1244
  • name: test
    num_bytes: 19149816
    num_examples: 1964
    download_size: 132925528
    dataset_size: 252764464
• config_name: gretel-safety-alignment
  features:
  • name: id
    dtype: int64
  • name: uuid
    dtype: string
  • name: prompt
    dtype: string
  • name: reasoning_20b
    dtype: string
  • name: response_20b
    dtype: string
  • name: reasoning_120b
    dtype: string
  • name: response_120b
    dtype: string
  • name: metadata
    struct:
    • name: judge_response_reasoning
      dtype: string
    • name: judge_response_score
      dtype: float64
    • name: judge_safe_response_reasoning
      dtype: string
    • name: judge_safe_response_score
      dtype: float64
    • name: persona
      dtype: string
    • name: response
      dtype: string
    • name: response_probability_of_harm
      dtype: float64
    • name: risk_category
      dtype: string
    • name: safe_response
      dtype: string
    • name: safe_response_probability_of_harm
      dtype: float64
    • name: sub_category
      dtype: string
    • name: tactic
      dtype: string
      splits:
  • name: train
    num_bytes: 51091205
    num_examples: 5994
  • name: validation
    num_bytes: 10111137
    num_examples: 1181
  • name: test
    num_bytes: 9857801
    num_examples: 1183
    download_size: 33554436
    dataset_size: 71060143
• config_name: malicious-tasks
  features:
  • name: id
    dtype: int64
  • name: uuid
    dtype: string
  • name: prompt
    dtype: string
  • name: reasoning_20b
    dtype: string
  • name: response_20b
    dtype: string
  • name: reasoning_120b
    dtype: string
  • name: response_120b
    dtype: string
  • name: metadata
    struct:
    • name: category
      dtype: string
    • name: severity
      dtype: string
    • name: subcategory
      dtype: string
      splits:
  • name: train
    num_bytes: 1255082
    num_examples: 225
    download_size: 627980
    dataset_size: 1255082
• config_name: redteam2k
  features:
  • name: id
    dtype: string
  • name: uuid
    dtype: string
  • name: prompt
    dtype: string
  • name: reasoning_20b
    dtype: string
  • name: response_20b
    dtype: string
  • name: reasoning_120b
    dtype: string
  • name: response_120b
    dtype: string
  • name: metadata
    struct:
    • name: from
      dtype: string
    • name: policy
      dtype: string
      splits:
  • name: train
    num_bytes: 17266707
    num_examples: 2000
    download_size: 9034588
    dataset_size: 17266707
    configs:
• config_name: aegis
  data_files:
  • split: train
    path: aegis/train-*
  • split: validation
    path: aegis/validation-*
  • split: test
    path: aegis/test-*
• config_name: gretel-safety-alignment
  data_files:
  • split: train
    path: gretel-safety-alignment/train-*
  • split: validation
    path: gretel-safety-alignment/validation-*
  • split: test
    path: gretel-safety-alignment/test-*
• config_name: malicious-tasks
  data_files:
  • split: train
    path: malicious-tasks/train-*
• config_name: redteam2k
  data_files:
  • split: train
    path: redteam2k/train-*
    task_categories:
• text-generation
• question-answering
  language:
• en
  tags:
• gpt-oss
• distillation
• reasoning
• ai-safety
  size_categories:
• 10K<n<100K

Nemotron Nano 2 Safety Distill — GPT-OSS

A distilled safety dataset produced using the Nemotron Nano 2 recipe with GPT-OSS-20B and GPT-OSS-120B as teacher models.

⚠️ Content Warning: This dataset includes potentially harmful prompts. Use responsibly for research purposes only.

Overview

This safety-focused distilled dataset was created by following the Nemotron Nano 2 safety recipe, adapted to use GPT-OSS-20B and GPT-OSS-120B as teacher models. Due to resource limitations, generated reasoning steps and responses were not filtered by guard models.

As of October 21, 2025, the dataset contains approximately 35,000 examples; the size may increase in future releases.

Data Sources

1. Aegis AI Content Safety Dataset v2.0 (Ghosh et al., 2025)

   • Paper: AEGIS2.0: A Diverse AI Safety Dataset and Risks Taxonomy for Alignment of LLM Guardrails
   • Dataset: nvidia/Aegis-AI-Content-Safety-Dataset-2.0

2. Gretel Safety Alignment Dataset v1 (Gretel AI, 2024)

   • Dataset: gretelai/gretel-safety-alignment-en-v1

3. HarmfulTasks Dataset (Hasan et al., 2024)

   • Paper: Pruning for Protection: Increasing Jailbreak Resistance in Aligned LLMs Without Fine-Tuning
   • Dataset: malicious_tasks_dataset.yaml

4. RedTeam-2K Dataset (Luo et al., 2024)

   • Paper: JailBreakV-28K: A Benchmark for Assessing the Robustness of Multi-modal Large Language Models against Jailbreak Attacks
   • Dataset: RedTeam-2K

Data Structure

This dataset comprises four subsets derived from the sources above:

• aegis: data points from the Aegis AI Content Safety Dataset v2.0
• gretel-safety-alignment: data points from the Gretel Safety Alignment Dataset v1
• malicious-tasks: data points from the HarmfulTasks Dataset
• redteam2k: data points from the RedTeam-2K Dataset

Each data point contains the following fields:

• id (string): Index within the subset
• uuid (string): Globally unique identifier across all subsets
• prompt (string): The input prompt that may contain harmful or unsafe content
• reasoning_20b (string): GPT-OSS-20B's reasoning steps
• response_20b (string): GPT-OSS-20B's response
• reasoning_120b (string): GPT-OSS-120B's reasoning steps
• response_120b (string): GPT-OSS-120B's response
• metadata (object): Additional context from the source dataset, such as harm category tags provided by the source dataset creators.

Generation Setup

python3.12 -m venv .venv
source .venv/bin/activate
pip install --pre vllm==0.10.1+gptoss \
  --extra-index-url https://wheels.vllm.ai/gpt-oss/ \
  --extra-index-url https://download.pytorch.org/whl/nightly/cu128 \
  --index-strategy unsafe-best-match

vllm serve openai/gpt-oss-20b
vllm serve openai/gpt-oss-120b

docker run -d \
  --gpus '"device=0,1"' \
  --ipc=host \
  --restart unless-stopped \
  -v /data/models/gpt-oss-120b:/models/gpt-oss-120b \
  -p 8802:8802 \
  --env VLLM_ATTENTION_BACKEND=TRITON_ATTN_VLLM_V1 \
  --env TORCH_CUDA_ARCH_LIST=8.6 \
  vllm/vllm-openai:gptoss \
  --served-model-name gptoss-120b \
  --tensor-parallel-size=2 \
  --model /models/gpt-oss-120b \
  --port 8802

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")

resp = client.chat.completions.create(
    model="openai/gpt-oss-20b",
    messages=[{"role": "user", "content": "Example query"}],
    temperature=1.0,
    top_p=1.0,
    max_tokens=16000,
    reasoning_effort="high"
)

print("Reasoning:", resp.choices[0].message.reasoning_content)
print("Response:", resp.choices[0].message.content)