dataset_info:
features:

• name: id
  dtype: string
• name: question
  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: 17186707
  num_examples: 2000
  download_size: 8967457
  dataset_size: 17186707
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
    license: mit
    task_categories:
• text-generation
• question-answering
  language:
• en
  tags:
• gpt-oss
• distillation
• reasoning
• ai-safety
  pretty_name: GPT-OSS Distilled RedTeam-2K
  size_categories:
• 1K<n<10K

GPT-OSS Distilled RedTeam-2K Dataset

This is a preliminary experimental subset of a larger dataset. For the full dataset and additional information, see: Nemotron Nano 2 Safety Distill — GPT-OSS .

⚠️ Content Warning: This dataset contains potentially harmful or policy-violating prompts (e.g., animal abuse, violence, privacy violations). The content includes sensitive safety-related queries and should be used responsibly for research purposes only.

Overview

This dataset contains safety knowledge distilled from the GPT-OSS-20B and GPT-OSS-120B models using high reasoning effort (MXFP4 quantized). It was created by processing all 2,000 queries from the RedTeam-2K dataset through both models to generate detailed reasoning processes and safety-aligned responses to potentially harmful prompts.

Source Dataset

This dataset is derived from the RedTeam-2K subset of the JailBreakV-28K dataset.

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

Dataset Structure

This dataset has following fields:

• id (string): Unique identifier for each example (0-1999)
• question (string): The original potentially harmful query from RedTeam-2K
• reasoning_20b (string): GPT-OSS-20B's internal reasoning steps
• response_20b (string): GPT-OSS-20B's response
• reasoning_120b (string): GPT-OSS-120B's internal reasoning steps
• response_120b (string): GPT-OSS-120B's response
• metadata (object): Additional information from the source dataset:
  • policy: Safety policy category (e.g., "Animal Abuse", "Violence", etc.)
  • from: The upstream source of the query

Dataset Usage Example

from datasets import load_dataset

dataset = load_dataset("Ericwang/gpt-oss-distilled-redteam2k")['train']
print(dataset[0])

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)