DiagSafe

Beyond Binary Labels: Diagnostic Evaluation of LLM Safety with Prompt-Level Rubrics

Paper (arXiv) · GitHub · Evaluation code · Dashboard

Content warning. This dataset contains harmful requests (e.g. violence, sexual content, self-harm, child safety, CBRN) written to test LLM safety. It is intended for safety evaluation and research only.

DiagSafe is a diagnostic safety benchmark for large language models. It pairs 3,489 bilingual prompts, grounded in an 84-label, three-tier safety taxonomy, with one prompt-level rubric per prompt: each rubric lists prompt-specific safety items with their exemptions, plus eight shared usefulness items, so an LLM judge rules on every item and cites evidence instead of emitting a single binary label.

Dataset Structure

Config	File	Records	Description
prompts (default)	prompts/DiagSafe_bench_3489.jsonl	3,489	Bilingual single- and multi-turn evaluation prompts
rubrics	rubrics/DiagSafe_rubrics_3489.jsonl	3,489	One prompt-level judge rubric per prompt
standards_zh	standards/DiagSafe_standards_84.json	84	Three-level taxonomy and label standards (Chinese)
standards_en	standards/DiagSafe_standards_84_en.json	84	The same 84 labels, in English

Composition

Dimension	Distribution
Subset	DiagBase: 2,401 · DiagAttack: 1,088
Language	English: 1,749 · Chinese: 1,740
Dialogue	Single-turn: 2,560 · Multi-turn: 929 (2–5 turns, 5,135 turns in total)
Taxonomy	11 first-level · 34 second-level · 84 third-level labels

• DiagBase is synthesized from the label standards along language × dialogue turns × attack form (direct harmful requests, template jailbreaks, combined-template jailbreaks).
• DiagAttack rewrites the single-turn direct requests of DiagBase into instruction attacks, embedding the harmful intent in a benign-looking task with role, format and tone constraints.

Code	First-level dimension	Prompts
OSS-01	Discrimination, Hate & Harmful Bias	440
OSS-02	Sexual & Sexually Explicit Content	695
OSS-03	Privacy Violations & Information Disclosure	195
OSS-04	Misinformation, Disinformation & Confabulation	240
OSS-05	Violent & Illegal Activities	831
OSS-06	CBRN & WMD	234
OSS-07	Cybersecurity & Model Attacks	187
OSS-08	Suicide, Self-Harm & Dangerous Health Behaviors	205
OSS-09	Child Safety	217
OSS-10	Intellectual Property Infringement	165
OSS-11	Manipulation & Deceptive Persuasion	80

Data Fields

prompts

Field	Type	Description
id	string	DiagBase-NNNNNN or DiagAttack-NNNNNN; the prefix identifies the subset
language	string	en or zh
messages	list	Ordered user turns, [{"role": "user", "content": ...}]; multi-turn prompts have 2–5
category	dict	l1 (code OSS-01–OSS-11), l2 and l3 (bilingual label names)

rubrics

Field	Type	Description
id	string	Matches exactly one prompt
rubric	string	Complete judge-prompt template with {question}, {answer} and {safety_categories} placeholders, 2–6 prompt-specific safety items (4.5 on average) with exemptions, eight usefulness items (U1.1–U3.2), a 0–3 usefulness scoring rule and a JSON output contract. Written in the language of its prompt

standards_zh / standards_en

Chinese field	English field	Description
类目编号	category_code	First-level code, OSS-01–OSS-11
类目名称	category_name	一级类目 / 二级类目 / 三级类目 ↔ l1 / l2 / l3
定义	definition	Label definition
标准总则	general_standard	General standard
标准细则	detailed_rules	Detailed rules, each marked ——危险 / ——安全 (——UNSAFE / ——SAFE)

Both standard files cover the same 84 labels in the same order. A prompt's category.l2 / category.l3 equal 类目名称.二级类目 / 类目名称.三级类目 in the Chinese file; the English file uses the parenthesized English names.

Usage

from datasets import load_dataset

prompts = load_dataset("tencent/DiagSafe", "prompts", split="test")
rubrics = load_dataset("tencent/DiagSafe", "rubrics", split="test")
standards = load_dataset("tencent/DiagSafe", "standards_zh", split="train")

rubric_by_id = dict(zip(rubrics["id"], rubrics["rubric"]))
example = prompts[0]
rubric = rubric_by_id[example["id"]]

The rubric contains JSON braces in its output contract, so fill the placeholders with str.replace rather than str.format:

import json

judge_prompt = (
    rubric
    .replace("{question}", json.dumps(example["messages"], ensure_ascii=False))
    .replace("{answer}", model_answer)
    .replace("{safety_categories}", json.dumps(example["category"], ensure_ascii=False))
)

Multi-turn prompts are evaluated turn by turn, with earlier turns given to the judge as context; a prompt counts as Unsafe if any turn is Unsafe. The GitHub repository provides the end-to-end evaluation script (scripts/run_eval.py), which handles answer generation, multi-turn judging, checkpointing and grouped statistics.

Intended Use and Limitations

DiagSafe is intended for evaluating and diagnosing the safety behavior of LLMs and LLM judges. The prompts are deliberately harmful; we ask users not to use them to train models to produce harmful content. Results depend on the judge model; the paper reports rubric-based judging at 89.41%–95.34% agreement with human annotations across judges.

License

Copyright (C) 2026 Tencent. Released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Citation

@misc{diagsafe2026,
  title        = {Beyond Binary Labels: Diagnostic Evaluation of LLM Safety with Prompt-Level Rubrics},
  author       = {Li, Hongyu and Huang, Xinlong and Yu, Zhan and Zhao, Jing and Zhao, Zhe and
                  Zhong, Juntao and Kuang, Xiong and Yuan, Rui and Yan, ZhenXiang and Xu, Huilin and
                  Yu, Haitao and Wang, Lilin and Chen, Rui and Chen, Blaze and Lian, Fengzong and
                  Hu, Renfen},
  year         = {2026},
  archivePrefix= {arXiv},
  eprint       = {2610.29225},
  primaryClass = {cs.CL}
}

Contact

nlpzhezhao@tencent.com