NovelPrompts

NovelPrompts is an English safety dataset (194 prompts + completions) built to test LLM judges' behaviour when evaluating the safety of prompts. The prompts are designed such that safety can only be assessed if you understand a novel concept: a recent event, a new word, or a new meaning of an existing word (e.g. slang). A concept is considered novel if it appeared after July 2025. This dataset is intended for evaluating LLM-as-judge safety evaluators, especially whether they can learn from context when evaluating novel concepts.

Annotators wrote safe and unsafe user requests that hinge on information that only became available after July 2025, chosen to investigate how LLMs with training cutoffs before this date can adapt to novel concepts. By design, prompt safety cannot be decided without understanding the novel concept. Context on each novel concept is also provided (e.g. "Bombakhalas: something that is crazy and like it's about to finish"), to investigate whether LLMs without knowledge of the novel concept can learn from context and correctly judge safety.

The 194 prompts contain references to five safety categories: SSH (self-harm / suicide), CSAM, MISI (mis/disinformation), Violence and Hate, and Sexual Content. We consider prompts to be unsafe if they mention a safety category in a way that is high-severity (e.g., inciting harm). If the reference is just neutral information content, or if the prompt does not mention any safety category, it is considered safe. Out of the 194 prompts, 61 are unsafe and 133 are safe.

For each prompt, a completion was generated by one of 5 frontier models: command-a-03-2025,DeepSeek-R1, gpt-4o-2024-05-13,command-r7b-12-2024, and claude-3.7-sonnet. The safety category and severity of each completion is also provided (as determined by expert annotators).

The dataset accompanies the paper Safety is Contextual, Judges are Not: Navigating the Rigid Priors of Evaluators.

Data collection

Human annotators at Alice, an AI data and safety company, wrote the novel-concept prompts and contexts, and produced the prompt-level safety annotations. The completion-level safety annotations were produced by internal Cohere annotators.

Columns

Column	Type	Description
prompt	string	The user prompt, written to lean on a novel concept whose meaning is only fully clear given prompt_context.
response	string	The model's response to the prompt.
estimator	string	The model that produced the response (one of the 5 listed above).
language	string	Always English.
prompt_context	string	A 1 to 2 sentence definition of the novel concept used in the prompt (e.g. "Bunker Buster: the topic went viral around June 18 2025…"). Always present.
prompt_category	string	Safety category the prompt falls under. NaN when the novel concept isn't in any safety category (e.g. benign slang).
prompt_severity	float	1.0 = unsafe prompt, 0.0 = the prompt touches a safety category but is low severity, NaN = the prompt does not mention a safety category. (So "safe" = 0.0 or NaN; 48 + 85 = 133.)
response_category	string	Same as prompt_category but evaluated on the model response.
response_severity	float	Same as prompt_severity but evaluated on the model response.
safety_assessment_without_context	string	yes if annotators can judge safety without knowing the context, no otherwise.
response_correct_interpretation	string	yes if the response appears to correctly interpret the novel concept given the context, no otherwise.
metadata	string	Python-dict-as-string with {prompt_id, language, estimator}.

Data loading

from datasets import load_dataset

ds = load_dataset("anissa218/novelprompts", split="test")

Or directly with pandas:

import pandas as pd

df = pd.read_csv("hf://datasets/anissa218/novelprompts/eval_novel_prompts_194_annotated.csv")

(For a private repo you'll need to be logged in via huggingface-cli login or have HF_TOKEN set.)

Citation

@article{alloula2026safetycontextualllmjudgesnot,
      title={Safety is Contextual, LLM-Judges Are Not: Navigating the Rigid Priors of Evaluators}, 
      author={Anissa Alloula and Federico Licini and Ava Batchkala and Seraphina Goldfarb-Tarrant},
      year={2026},
      eprint={2606.07874},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2606.07874}, 
}