andyrdt/gpt-oss-20b-rollouts

configs:

• config_name: BeaverTails
  data_files:
  • split: 330k_train
    path: BeaverTails/330k_train.jsonl
• config_name: HarmBench
  data_files:
  • split: copyright_train
    path: HarmBench/copyright_train.jsonl
  • split: standard_train
    path: HarmBench/standard_train.jsonl
• config_name: HarmfulGeneration-HarmBench
  data_files:
  • split: test
    path: HarmfulGeneration-HarmBench/test.jsonl
• config_name: JBB-Behaviors
  data_files:
  • split: judge_comparison_test
    path: JBB-Behaviors/judge_comparison_test.jsonl
• config_name: MATH-500
  data_files:
  • split: test
    path: MATH-500/test.jsonl
• config_name: NuminaMath-CoT
  data_files:
  • split: train
    path: NuminaMath-CoT/train.jsonl
• config_name: StrongREJECT
  data_files:
  • split: train
    path: StrongREJECT/train.jsonl
• config_name: WildChat-1M
  data_files:
  • split: train
    path: WildChat-1M/train.jsonl
• config_name: XSTest
  data_files:
  • split: train
    path: XSTest/train.jsonl
• config_name: apps
  data_files:
  • split: train
    path: apps/train.jsonl
• config_name: apps-materialized
  data_files:
  • split: train
    path: apps-materialized/train.jsonl
• config_name: circuit-breakers-dataset
  data_files:
  • split: train
    path: circuit-breakers-dataset/train.jsonl
• config_name: combined
  data_files:
  • split: all_shuffled
    path: combined/all_shuffled.jsonl
• config_name: gandalf_ignore_instructions
  data_files:
  • split: train
    path: gandalf_ignore_instructions/train.jsonl
• config_name: gsm8k
  data_files:
  • split: main_train
    path: gsm8k/main_train.jsonl
• config_name: mbpp
  data_files:
  • split: full_test
    path: mbpp/full_test.jsonl
  • split: full_train
    path: mbpp/full_train.jsonl
• config_name: mmlu
  data_files:
  • split: all_test
    path: mmlu/all_test.jsonl
• config_name: oasst1
  data_files:
  • split: train
    path: oasst1/train.jsonl
• config_name: openai_humaneval
  data_files:
  • split: test
    path: openai_humaneval/test.jsonl
• config_name: or-bench
  data_files:
  • split: or_bench_80k_train
    path: or-bench/or-bench-80k_train.jsonl
  • split: or_bench_hard_1k_train
    path: or-bench/or-bench-hard-1k_train.jsonl
  • split: or_bench_toxic_train
    path: or-bench/or-bench-toxic_train.jsonl
• config_name: ultrachat_200k
  data_files:
  • split: train_sft
    path: ultrachat_200k/train_sft.jsonl
    license: mit

GPT-OSS-20B Rollouts

Generated rollouts from GPT-OSS-20B with parsed Harmony channels (assistant thinking/final).

• Schema: user_content, system_reasoning_effort, assistant_thinking, assistant_content.
• Loading example: load_dataset("andyrdt/gpt-oss-20b-rollouts", "HarmBench", split="standard_train").

Notes

• This repository uses manual configuration to expose both subset (config) and split dropdowns in the viewer.

Safety and jailbreak

• HarmBench: Safety prompts, standard subset
  • Splits: standard_train, copyright_train
• HarmfulGeneration-HarmBench: Harmful generation evaluation prompts
  • Splits: test
• StrongREJECT: Refusal/defense behaviors
  • Splits: train
• JBB-Behaviors: Jailbreak evaluation prompts
  • Splits: judge_comparison_test
• circuit-breakers-dataset: Adversarial prompts to trip safety
  • Splits: train
• gandalf_ignore_instructions: Instruction-ignoring adversarial tasks
  • Splits: train
• BeaverTails: Safety alignment dataset
  • Splits: 330k_train
• or-bench: Over-refusal benchmark (80k seemingly toxic prompts)
  • Splits: or-bench-80k_train, or-bench-hard-1k_train, or-bench-toxic_train
• XSTest: Exaggerated safety behaviors test suite
  • Splits: train

Math and reasoning

• gsm8k: Grade school math problems
  • Splits: main_train
• MATH-500: Math problems for evaluation
  • Splits: test
• mmlu: Multiple-choice reasoning benchmark (all subjects)
  • Splits: all_test
• NuminaMath-CoT: Math competition problems
  • Splits: train

Chat and instruction

• WildChat-1M: Diverse open-domain chats
  • Splits: train
• ultrachat_200k: Large-scale assistant conversations
  • Splits: train_sft
• oasst1: OpenAssistant curated conversations
  • Splits: train

Code / programming

• openai_humaneval: Code generation evaluation (HumanEval)
  • Splits: test
• mbpp: MBPP training set
  • Splits: full_train, full_test
• apps: APPS coding problems
  • Splits: train

How to load

from datasets import load_dataset
ds = load_dataset("andyrdt/gpt-oss-20b-rollouts", "HarmBench", split="standard_train")
# e.g., BeaverTails 330k split:
ds_bt = load_dataset("andyrdt/gpt-oss-20b-rollouts", "BeaverTails", split="330k_train")

Prompt formatting

from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("openai/gpt-oss-20b", trust_remote_code=True)

conversation = [
    {"role": "user", "content": user_content},
    {"role": "assistant", "content": assistant_content, "thinking": assistant_thinking},
]

formatted_conversation = tok.apply_chat_template(
    conversation,
    reasoning_effort=system_reasoning_effort,
    add_generation_prompt=False,
    tokenize=False,
)