Welcome to Inspect Evals, a repository of community contributed LLM evaluations for Inspect AI. Inspect Evals was created in collaboration by the UK AISI, Arcadia Impact, and the Vector Institute.

📚 Documentation

Community contributions are welcome and encouraged! Please see the Contributor Guide for details on submitting new evaluations.

First, install the necessary dependencies with:

uv sync

Some of the evaluations require additional dependencies. If your eval needs extra dependencies, instructions for installing them are provided in the list of evals subsection (or the README for that evaluation). For example, to install the dependencies of SWE-Bench evaluation you should run:

uv sync --extra swe_bench

Once you have a model configured, you can run evaluations for it with:

uv run inspect eval inspect_evals/gpqa_diamond --model openai/gpt-4o

If you don't want to specify the --model each time you run an evaluation, create a .env configuration file in your working directory that defines the INSPECT_EVAL_MODEL environment variable along with your API key. For example:

INSPECT_EVAL_MODEL=openai/gpt-4o
OPENAI_API_KEY=<openai-api-key>

Inspect supports many model providers including OpenAI, Anthropic, Google, Mistral, Azure AI, AWS Bedrock, Together AI, Groq, Hugging Face, vLLM, Ollama, and more. See the Model Providers documentation for additional details.

You might also be able to use a newer version of pip (25.1+) to install the project via pip install --group dev .' or pip install --group dev '.[swe_bench]'`. However this is not officially supported.

For details on building the documentation, see the documentation guide.

For information on running tests and CI toggles, see the Technical Contribution Guide in CONTRIBUTING.md.

We recommend having at least 35 GB of free disk space for Inspect Evals: the full installation takes about 10 GB and you'll also need some space for uv cache and datasets cache (most are small, but some take 13 GB such as MMIU).

Running some evals (e.g., CyBench, GDM capabilities evals) may require extra space beyond this because they pull Docker images. We recommend having at least 65 GB of extra space for running evals that have Dockerfiles in their file tree (though you might get away with less space) on top of the 35 GB suggestion above.

In total, you should be comfortable running evals with 100 GB of free space. If you end up running of out space while having 100+ GB of free space available, please let us know — this might be a bug.

The amount of memory needed for an eval varies significantly with the eval. You'll be able to run most evals with only 0.5 GB of free RAM. However, some evals with larger datasets require 2-3 GB or more. And some evals that use Docker (e.g., some GDM capabilities evals) require up to 32 GB of RAM.

• APPS: Automated Programming Progress Standard

  APPS is a dataset for evaluating model performance on Python programming tasks across three difficulty levels consisting of 1,000 at introductory, 3,000 at interview, and 1,000 at competition level. The dataset consists of an additional 5,000 training samples, for a total of 10,000 total samples. We evaluate on questions from the test split, which consists of programming problems commonly found in coding interviews. _{^{Contributed by: @camtice}}

  uv run inspect eval inspect_evals/apps

• AgentBench: Evaluate LLMs as Agents

  A benchmark designed to evaluate LLMs as Agents _{^{Contributed by: @Felhof, @hannagabor, @shaheenahmedc}}

  uv run inspect eval inspect_evals/agent_bench_os

• BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions

  Python coding benchmark with 1,140 diverse questions drawing on numerous python libraries. _{^{Contributed by: @tim-hua-01}}

  uv run inspect eval inspect_evals/bigcodebench

• CORE-Bench

  Evaluate how well an LLM Agent is at computationally reproducing the results of a set of scientific papers. _{^{Contributed by: @enerrio}}

  uv run inspect eval inspect_evals/core_bench

• ClassEval: A Manually-Crafted Benchmark for Evaluating LLMs on Class-level Code Generation

  Evaluates LLMs on class-level code generation with 100 tasks constructed over 500 person-hours. The study shows that LLMs perform worse on class-level tasks compared to method-level tasks. _{^{Contributed by: @zhenningdavidliu}}

  uv run inspect eval inspect_evals/class_eval

• DS-1000: A Natural and Reliable Benchmark for Data Science Code Generation

  Code generation benchmark with a thousand data science problems spanning seven Python libraries. _{^{Contributed by: @bienehito}}

  uv run inspect eval inspect_evals/ds1000

• HumanEval: Python Function Generation from Instructions

  Assesses how accurately language models can write correct Python functions based solely on natural-language instructions provided as docstrings. _{^{Contributed by: @adil-a}}

  uv run inspect eval inspect_evals/humaneval

• MBPP: Basic Python Coding Challenges

  Measures the ability of language models to generate short Python programs from simple natural-language descriptions, testing basic coding proficiency. _{^{Contributed by: @jddantes}}

  uv run inspect eval inspect_evals/mbpp

• MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering

  Machine learning tasks drawn from 75 Kaggle competitions. _{^{Contributed by: @samm393}}

  uv run inspect eval inspect_evals/mle_bench
  uv run inspect eval inspect_evals/mle_bench_full
  uv run inspect eval inspect_evals/mle_bench_lite

• SWE-bench Verified: Resolving Real-World GitHub Issues

  Evaluates AI's ability to resolve genuine software engineering issues sourced from 12 popular Python GitHub repositories, reflecting realistic coding and debugging scenarios. _{^{Contributed by: @max-kaufmann}}

  uv run inspect eval inspect_evals/swe_bench
  uv run inspect eval inspect_evals/swe_bench_verified_mini

• SciCode: A Research Coding Benchmark Curated by Scientists

  SciCode tests the ability of language models to generate code to solve scientific research problems. It assesses models on 65 problems from mathematics, physics, chemistry, biology, and materials science. _{^{Contributed by: @xantheocracy}}

  uv run inspect eval inspect_evals/scicode

• USACO: USA Computing Olympiad

  Evaluates language model performance on difficult Olympiad programming problems across four difficulty levels. _{^{Contributed by: @danwilhelm}}

  uv run inspect eval inspect_evals/usaco

• AssistantBench: Can Web Agents Solve Realistic and Time-Consuming Tasks?

  Tests whether AI agents can perform real-world time-consuming tasks on the web. _{^{Contributed by: @nlpet, @caspardh}}

  uv run inspect eval inspect_evals/assistant_bench_closed_book_zero_shot
  uv run inspect eval inspect_evals/assistant_bench_closed_book_one_shot
  uv run inspect eval inspect_evals/assistant_bench_web_search_zero_shot
  uv run inspect eval inspect_evals/assistant_bench_web_search_one_shot
  uv run inspect eval inspect_evals/assistant_bench_web_browser

• BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents

  A benchmark for evaluating agents' ability to browse the web. The dataset consists of challenging questions that generally require web-access to answer correctly. _{^{Contributed by: @AnselmC}}

  uv run inspect eval inspect_evals/browse_comp

• GAIA: A Benchmark for General AI Assistants

  Proposes real-world questions that require a set of fundamental abilities such as reasoning, multi-modality handling, web browsing, and generally tool-use proficiency. GAIA questions are conceptually simple for humans yet challenging for most advanced AIs. _{^{Contributed by: @max-kaufmann}}

  uv run inspect eval inspect_evals/gaia
  uv run inspect eval inspect_evals/gaia_level1
  uv run inspect eval inspect_evals/gaia_level2
  uv run inspect eval inspect_evals/gaia_level3

• Mind2Web: Towards a Generalist Agent for the Web

  A dataset for developing and evaluating generalist agents for the web that can follow language instructions to complete complex tasks on any website. _{^{Contributed by: @dr3s}}

  uv run inspect eval inspect_evals/mind2web

• OSWorld: Multimodal Computer Interaction Tasks

  Tests AI agents' ability to perform realistic, open-ended tasks within simulated computer environments, requiring complex interaction across multiple input modalities. _{^{Contributed by: @epatey}}

  uv run inspect eval inspect_evals/osworld
  uv run inspect eval inspect_evals/osworld_small

• Sycophancy Eval

  Evaluate sycophancy of language models across a variety of free-form text-generation tasks. _{^{Contributed by: @alexdzm}}

  uv run inspect eval inspect_evals/sycophancy

• CYBERSECEVAL 3: Advancing the Evaluation of Cybersecurity Risks and Capabilities in Large Language Models

  Evaluates Large Language Models for cybersecurity risk to third parties, application developers and end users. _{^{Contributed by: @onionymous}}

  uv run inspect eval inspect_evals/cyse3_visual_prompt_injection

• Catastrophic Cyber Capabilities Benchmark (3CB): Robustly Evaluating LLM Agent Cyber Offense Capabilities

  A benchmark for evaluating the capabilities of LLM agents in cyber offense. _{^{Contributed by: @derpyplops, @ralphptorres, @Scott-Simmons}}

  uv run inspect eval inspect_evals/threecb

• Cybench: Capture-The-Flag Cybersecurity Challenges

  Tests language models on cybersecurity skills using 40 practical, professional-level challenges taken from cybersecurity competitions, designed to cover various difficulty levels and security concepts. _{^{Contributed by: @sinman-aisi, @sam-deverett-dsit, @kola-aisi, @pgiav}}

  uv run inspect eval inspect_evals/cybench

• CyberMetric: A Benchmark Dataset based on Retrieval-Augmented Generation for Evaluating LLMs in Cybersecurity Knowledge

  Datasets containing 80, 500, 2000 and 10000 multiple-choice questions, designed to evaluate understanding across nine domains within cybersecurity _{^{Contributed by: @neilshaabi}}

  uv run inspect eval inspect_evals/cybermetric_80
  uv run inspect eval inspect_evals/cybermetric_500
  uv run inspect eval inspect_evals/cybermetric_2000
  uv run inspect eval inspect_evals/cybermetric_10000

• CyberSecEval_2: Cybersecurity Risk and Vulnerability Evaluation

  Assesses language models for cybersecurity risks, specifically testing their potential to misuse programming interpreters, vulnerability to malicious prompt injections, and capability to exploit known software vulnerabilities. _{^{Contributed by: @its-emile}}

  uv run inspect eval inspect_evals/cyse2_interpreter_abuse
  uv run inspect eval inspect_evals/cyse2_prompt_injection
  uv run inspect eval inspect_evals/cyse2_vulnerability_exploit

• GDM Dangerous Capabilities: Capture the Flag

  CTF challenges covering web app vulnerabilities, off-the-shelf exploits, databases, Linux privilege escalation, password cracking and spraying. Demonstrates tool use and sandboxing untrusted model code. _{^{Contributed by: @XkunW}}

  uv run inspect eval inspect_evals/gdm_in_house_ctf

• InterCode: Security and Coding Capture-the-Flag Challenges

  Tests AI's ability in coding, cryptography, reverse engineering, and vulnerability identification through practical capture-the-flag (CTF) cybersecurity scenarios. _{^{Contributed by: @jjallaire}}

  uv run inspect eval inspect_evals/gdm_intercode_ctf

• SEvenLLM: A benchmark to elicit, and improve cybersecurity incident analysis and response abilities in LLMs for Security Events.

  Designed for analyzing cybersecurity incidents, which is comprised of two primary task categories: understanding and generation, with a further breakdown into 28 subcategories of tasks. _{^{Contributed by: @kingroryg}}

  uv run inspect eval inspect_evals/sevenllm_mcq_zh
  uv run inspect eval inspect_evals/sevenllm_mcq_en
  uv run inspect eval inspect_evals/sevenllm_qa_zh
  uv run inspect eval inspect_evals/sevenllm_qa_en

• SecQA: A Concise Question-Answering Dataset for Evaluating Large Language Models in Computer Security

  "Security Question Answering" dataset to assess LLMs' understanding and application of security principles. SecQA has "v1" and "v2" datasets of multiple-choice questions that aim to provide two levels of cybersecurity evaluation criteria. The questions were generated by GPT-4 based on the "Computer Systems Security: Planning for Success" textbook and vetted by humans. _{^{Contributed by: @matthewreed26}}

  uv run inspect eval inspect_evals/sec_qa_v1
  uv run inspect eval inspect_evals/sec_qa_v1_5_shot
  uv run inspect eval inspect_evals/sec_qa_v2
  uv run inspect eval inspect_evals/sec_qa_v2_5_shot

• AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions

  Evaluating abstention across 20 diverse datasets, including questions with unknown answers, underspecification, false premises, subjective interpretations, and outdated information. _{^{Contributed by: @jeqcho}}

  uv run inspect eval inspect_evals/abstention_bench

• AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents

  Assesses whether AI agents can be hijacked by malicious third parties using prompt injections in simple environments such as a workspace or travel booking app. _{^{Contributed by: @eric-aisi}}

  uv run inspect eval inspect_evals/agentdojo

• AgentHarm: Harmfulness Potential in AI Agents

  Assesses whether AI agents might engage in harmful activities by testing their responses to malicious prompts in areas like cybercrime, harassment, and fraud, aiming to ensure safe behavior. _{^{Contributed by: @alexandrasouly-aisi, @eric-aisi, @max-andr, @xanderdavies}}

  uv run inspect eval inspect_evals/agentharm
  uv run inspect eval inspect_evals/agentharm_benign

• LAB-Bench: Measuring Capabilities of Language Models for Biology Research

  Tests LLMs and LLM-augmented agents abilities to answer questions on scientific research workflows in domains like chemistry, biology, materials science, as well as more general science tasks _{^{Contributed by: @matthewreed26}}

  uv run inspect eval inspect_evals/lab_bench_litqa
  uv run inspect eval inspect_evals/lab_bench_suppqa
  uv run inspect eval inspect_evals/lab_bench_figqa
  uv run inspect eval inspect_evals/lab_bench_tableqa
  uv run inspect eval inspect_evals/lab_bench_dbqa
  uv run inspect eval inspect_evals/lab_bench_protocolqa
  uv run inspect eval inspect_evals/lab_bench_seqqa
  uv run inspect eval inspect_evals/lab_bench_cloning_scenarios

• MASK: Disentangling Honesty from Accuracy in AI Systems

  Evaluates honesty in large language models by testing whether they contradict their own beliefs when pressured to lie. _{^{Contributed by: @Scott-Simmons}}

  uv run inspect eval inspect_evals/mask

• Make Me Pay

  Evaluates an AI models' susceptibility to social engineering attacks by testing whether a "con-artist" model can persuade a "mark" model to donate money through manipulation and persuasion tactics. _{^{Contributed by: @PranshuSrivastava}}

  uv run inspect eval inspect_evals/make_me_pay

• StereoSet: Measuring stereotypical bias in pretrained language models

  A dataset that measures stereotype bias in language models across gender, race, religion, and profession domains. Models choose between stereotype, anti-stereotype, and unrelated completions to sentences. _{^{Contributed by: @Xodarap}}

  uv run inspect eval inspect_evals/stereoset

• StrongREJECT: Measuring LLM susceptibility to jailbreak attacks

  A benchmark that evaluates the susceptibility of LLMs to various jailbreak attacks. _{^{Contributed by: @viknat}}

  uv run inspect eval inspect_evals/strong_reject

• WMDP: Measuring and Reducing Malicious Use With Unlearning

  A dataset of 3,668 multiple-choice questions developed by a consortium of academics and technical consultants that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. _{^{Contributed by: @alexandraabbas}}

  uv run inspect eval inspect_evals/wmdp_bio
  uv run inspect eval inspect_evals/wmdp_chem
  uv run inspect eval inspect_evals/wmdp_cyber

• AIME 2024: Problems from the American Invitational Mathematics Examination

  A benchmark for evaluating AI's ability to solve challenging mathematics problems from AIME - a prestigious high school mathematics competition. _{^{Contributed by: @tamazgadaev}}

  uv run inspect eval inspect_evals/aime2024

• GSM8K: Grade School Math Word Problems

  Measures how effectively language models solve realistic, linguistically rich math word problems suitable for grade-school-level mathematics. _{^{Contributed by: @jjallaire}}

  uv run inspect eval inspect_evals/gsm8k

• MATH: Measuring Mathematical Problem Solving

  Dataset of 12,500 challenging competition mathematics problems. Demonstrates fewshot prompting and custom scorers. NOTE: The dataset has been taken down due to a DMCA notice from The Art of Problem Solving. _{^{Contributed by: @xeon27}}

  uv run inspect eval inspect_evals/math

• MGSM: Multilingual Grade School Math

  Extends the original GSM8K dataset by translating 250 of its problems into 10 typologically diverse languages. _{^{Contributed by: @manifoldhiker}}

  uv run inspect eval inspect_evals/mgsm

• MathVista: Visual Math Problem-Solving

  Tests AI models on math problems that involve interpreting visual elements like diagrams and charts, requiring detailed visual comprehension and logical reasoning. _{^{Contributed by: @ShivMunagala}}

  uv run inspect eval inspect_evals/mathvista

• ARC: AI2 Reasoning Challenge

  Dataset of natural, grade-school science multiple-choice questions (authored for human tests). _{^{Contributed by: @jjallaire}}

  uv run inspect eval inspect_evals/arc_easy
  uv run inspect eval inspect_evals/arc_challenge

• BBH: Challenging BIG-Bench Tasks

  Tests AI models on a suite of 23 challenging BIG-Bench tasks that previously proved difficult even for advanced language models to solve. _{^{Contributed by: @JoschkaCBraun}}

  uv run inspect eval inspect_evals/bbh

• BIG-Bench Extra Hard

  A reasoning capability dataset that replaces each task in BIG-Bench-Hard with a novel task that probes a similar reasoning capability but exhibits significantly increased difficulty. _{^{Contributed by: @jeqcho}}

  uv run inspect eval inspect_evals/bbeh
  uv run inspect eval inspect_evals/bbeh_mini

• BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions

  Reading comprehension dataset that queries for complex, non-factoid information, and require difficult entailment-like inference to solve. _{^{Contributed by: @seddy-aisi}}

  uv run inspect eval inspect_evals/boolq

• DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs

  Evaluates reading comprehension where models must resolve references in a question, perhaps to multiple input positions, and perform discrete operations over them (such as addition, counting, or sorting). _{^{Contributed by: @xeon27}}

  uv run inspect eval inspect_evals/drop

• HellaSwag: Commonsense Event Continuation

  Tests models' commonsense reasoning abilities by asking them to select the most likely next step or continuation for a given everyday situation. _{^{Contributed by: @jjallaire}}

  uv run inspect eval inspect_evals/hellaswag

• IFEval: Instruction-Following Evaluation

  Evaluates how well language models can strictly follow detailed instructions, such as writing responses with specific word counts or including required keywords. _{^{Contributed by: @adil-a}}

  uv run inspect eval inspect_evals/ifeval

• LingOly

  Two linguistics reasoning benchmarks: LingOly (Linguistic Olympiad questions) is a benchmark utilising low resource languages. LingOly-TOO (Linguistic Olympiad questions with Templatised Orthographic Obfuscation) is a benchmark designed to counteract answering without reasoning. _{^{Contributed by: @am-bean, @jkhouja}}

  uv run inspect eval inspect_evals/lingoly
  uv run inspect eval inspect_evals/lingoly_too

• MMMU: Multimodal College-Level Understanding and Reasoning

  Assesses multimodal AI models on challenging college-level questions covering multiple academic subjects, requiring detailed visual interpretation, in-depth reasoning, and both multiple-choice and open-ended answering abilities. _{^{Contributed by: @shaheenahmedc}}

  uv run inspect eval inspect_evals/mmmu_multiple_choice
  uv run inspect eval inspect_evals/mmmu_open

• MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning

  Evaluating models on multistep soft reasoning tasks in the form of free text narratives. _{^{Contributed by: @farrelmahaztra}}

  uv run inspect eval inspect_evals/musr

• Needle in a Haystack (NIAH): In-Context Retrieval Benchmark for Long Context LLMs

  NIAH evaluates in-context retrieval ability of long context LLMs by testing a model's ability to extract factual information from long-context inputs. _{^{Contributed by: @owenparsons}}

  uv run inspect eval inspect_evals/niah

• PAWS: Paraphrase Adversaries from Word Scrambling

  Evaluating models on the task of paraphrase detection by providing pairs of sentences that are either paraphrases or not. _{^{Contributed by: @meltemkenis}}

  uv run inspect eval inspect_evals/paws

• PIQA: Physical Commonsense Reasoning Test

  Measures the model's ability to apply practical, everyday commonsense reasoning about physical objects and scenarios through simple decision-making questions. _{^{Contributed by: @seddy-aisi}}

  uv run inspect eval inspect_evals/piqa

• RACE-H: A benchmark for testing reading comprehension and reasoning abilities of neural models

  Reading comprehension tasks collected from the English exams for middle and high school Chinese students in the age range between 12 to 18. _{^{Contributed by: @mdrpanwar}}

  uv run inspect eval inspect_evals/race_h

• SQuAD: A Reading Comprehension Benchmark requiring reasoning over Wikipedia articles

  Set of 100,000+ questions posed by crowdworkers on a set of Wikipedia articles, where the answer to each question is a segment of text from the corresponding reading passage. _{^{Contributed by: @tknasir}}

  uv run inspect eval inspect_evals/squad

• WINOGRANDE: An Adversarial Winograd Schema Challenge at Scale

  Set of 273 expert-crafted pronoun resolution problems originally designed to be unsolvable for statistical models that rely on selectional preferences or word associations. _{^{Contributed by: @xeon27}}

  uv run inspect eval inspect_evals/winogrande

• ∞Bench: Extending Long Context Evaluation Beyond 100K Tokens

  LLM benchmark featuring an average data length surpassing 100K tokens. Comprises synthetic and realistic tasks spanning diverse domains in English and Chinese. _{^{Contributed by: @celiawaggoner}}

  uv run inspect eval inspect_evals/infinite_bench_code_debug
  uv run inspect eval inspect_evals/infinite_bench_code_run
  uv run inspect eval inspect_evals/infinite_bench_kv_retrieval
  uv run inspect eval inspect_evals/infinite_bench_longbook_choice_eng
  uv run inspect eval inspect_evals/infinite_bench_longdialogue_qa_eng
  uv run inspect eval inspect_evals/infinite_bench_math_calc
  uv run inspect eval inspect_evals/infinite_bench_math_find
  uv run inspect eval inspect_evals/infinite_bench_number_string
  uv run inspect eval inspect_evals/infinite_bench_passkey

• AGIEval: A Human-Centric Benchmark for Evaluating Foundation Models

  AGIEval is a human-centric benchmark specifically designed to evaluate the general abilities of foundation models in tasks pertinent to human cognition and problem-solving. _{^{Contributed by: @bouromain}}

  uv run inspect eval inspect_evals/agie_aqua_rat
  uv run inspect eval inspect_evals/agie_logiqa_en
  uv run inspect eval inspect_evals/agie_lsat_ar
  uv run inspect eval inspect_evals/agie_lsat_lr
  uv run inspect eval inspect_evals/agie_lsat_rc
  uv run inspect eval inspect_evals/agie_math
  uv run inspect eval inspect_evals/agie_sat_en
  uv run inspect eval inspect_evals/agie_sat_en_without_passage
  uv run inspect eval inspect_evals/agie_sat_math

• AIR Bench: AI Risk Benchmark

  A safety benchmark evaluating language models against risk categories derived from government regulations and company policies. _{^{Contributed by: @l1990790120}}

  uv run inspect eval inspect_evals/air_bench

• ChemBench: Are large language models superhuman chemists?

  ChemBench is designed to reveal limitations of current frontier models for use in the chemical sciences. It consists of 2786 question-answer pairs compiled from diverse sources. Our corpus measures reasoning, knowledge and intuition across a large fraction of the topics taught in undergraduate and graduate chemistry curricula. It can be used to evaluate any system that can return text (i.e., including tool-augmented systems). _{^{Contributed by: @Esther-Guo}}

  uv run inspect eval inspect_evals/chembench

• CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge

  Evaluates an AI model's ability to correctly answer everyday questions that rely on basic commonsense knowledge and understanding of the world. _{^{Contributed by: @jjallaire}}

  uv run inspect eval inspect_evals/commonsense_qa

• GPQA: Graduate-Level STEM Knowledge Challenge

  Contains challenging multiple-choice questions created by domain experts in biology, physics, and chemistry, designed to test advanced scientific understanding beyond basic internet searches. Experts at PhD level in the corresponding domains reach 65% accuracy. _{^{Contributed by: @jjallaire}}

  uv run inspect eval inspect_evals/gpqa_diamond

• HealthBench: Evaluating Large Language Models Towards Improved Human Health

  A comprehensive evaluation benchmark designed to assess language models' medical capabilities across a wide range of healthcare scenarios. _{^{Contributed by: @retroam}}

  uv run inspect eval inspect_evals/healthbench
  uv run inspect eval inspect_evals/healthbench_hard
  uv run inspect eval inspect_evals/healthbench_consensus

• Humanity's Last Exam

  Humanity's Last Exam (HLE) is a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. Humanity's Last Exam consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. _{^{Contributed by: @SasankYadati}}

  uv run inspect eval inspect_evals/hle

• LiveBench: A Challenging, Contamination-Free LLM Benchmark

  LiveBench is a benchmark designed with test set contamination and objective evaluation in mind by releasing new questions regularly, as well as having questions based on recently-released datasets. Each question has verifiable, objective ground-truth answers, allowing hard questions to be scored accurately and automatically, without the use of an LLM judge. _{^{Contributed by: @anaoaktree}}

  uv run inspect eval inspect_evals/livebench

• MMLU-Pro: Advanced Multitask Knowledge and Reasoning Evaluation

  An advanced benchmark that tests both broad knowledge and reasoning capabilities across many subjects, featuring challenging questions and multiple-choice answers with increased difficulty and complexity. _{^{Contributed by: @xeon27}}

  uv run inspect eval inspect_evals/mmlu_pro

• MMLU: Measuring Massive Multitask Language Understanding

  Evaluate models on 57 tasks including elementary mathematics, US history, computer science, law, and more. _{^{Contributed by: @jjallaire, @domdomegg}}

  uv run inspect eval inspect_evals/mmlu_0_shot
  uv run inspect eval inspect_evals/mmlu_5_shot

• MedQA: Medical exam Q&A benchmark

  A Q&A benchmark with questions collected from professional medical board exams. Only includes the English subset of the dataset (which also contains Mandarin Chinese and Taiwanese questions). _{^{Contributed by: @bunny-baxter, @JasonBenn}}

  uv run inspect eval inspect_evals/medqa

• O-NET: A high-school student knowledge test

  Questions and answers from the Ordinary National Educational Test (O-NET), administered annually by the National Institute of Educational Testing Service to Matthayom 6 (Grade 12 / ISCED 3) students in Thailand. The exam contains six subjects: English language, math, science, social knowledge, and Thai language. There are questions with multiple-choice and true/false answers. Questions can be in either English or Thai. _{^{Contributed by: @bact}}

  uv run inspect eval inspect_evals/onet_m6

• Pre-Flight: Aviation Operations Knowledge Evaluation

  Tests model understanding of aviation regulations including ICAO annexes, flight dispatch rules, pilot procedures, and airport ground operations safety protocols. _{^{Contributed by: @alexbrooker}}

  uv run inspect eval inspect_evals/pre_flight

• PubMedQA: A Dataset for Biomedical Research Question Answering

  Biomedical question answering (QA) dataset collected from PubMed abstracts. _{^{Contributed by: @MattFisher}}

  uv run inspect eval inspect_evals/pubmedqa

• SOS BENCH: Benchmarking Safety Alignment on Scientific Knowledge

  A regulation-grounded, hazard-focused benchmark encompassing six high-risk scientific domains: chemistry, biology, medicine, pharmacology, physics, and psychology. The benchmark comprises 3,000 prompts derived from real-world regulations and laws, systematically expanded via an LLM-assisted evolutionary pipeline that introduces diverse, realistic misuse scenarios (e.g., detailed explosive synthesis instructions involving advanced chemical formulas). _{^{Contributed by: @Esther-Guo}}

  uv run inspect eval inspect_evals/sosbench

• SciKnowEval: Evaluating Multi-level Scientific Knowledge of Large Language Models

  The Scientific Knowledge Evaluation benchmark is inspired by the profound principles outlined in the “Doctrine of the Mean” from ancient Chinese philosophy. This benchmark is designed to assess LLMs based on their proficiency in Studying Extensively, Enquiring Earnestly, Thinking Profoundly, Discerning Clearly, and Practicing Assiduously. Each of these dimensions offers a unique perspective on evaluating the capabilities of LLMs in handling scientific knowledge. _{^{Contributed by: @Esther-Guo}}

  uv run inspect eval inspect_evals/sciknoweval

• SimpleQA/SimpleQA Verified: Measuring short-form factuality in large language models

  A benchmark that evaluates the ability of language models to answer short, fact-seeking questions. _{^{Contributed by: @osc245, @jeqcho}}

  uv run inspect eval inspect_evals/simpleqa
  uv run inspect eval inspect_evals/simpleqa_verified

• TruthfulQA: Measuring How Models Mimic Human Falsehoods

  Measure whether a language model is truthful in generating answers to questions using questions that some humans would answer falsely due to a false belief or misconception. _{^{Contributed by: @seddy-aisi}}

  uv run inspect eval inspect_evals/truthfulqa

• XSTest: A benchmark for identifying exaggerated safety behaviours in LLM's

  Dataset with 250 safe prompts across ten prompt types that well-calibrated models should not refuse, and 200 unsafe prompts as contrasts that models, for most applications, should refuse. _{^{Contributed by: @NelsonG-C}}

  uv run inspect eval inspect_evals/xstest

• Agentic Misalignment: How LLMs could be insider threats

  Eliciting unethical behaviour (most famously blackmail) in response to a fictional company-assistant scenario where the model is faced with replacement. _{^{Contributed by: @bmillwood-aisi}}

  uv run inspect eval inspect_evals/agentic_misalignment

• GDM Dangerous Capabilities: Self-reasoning

  Test AI's ability to reason about its environment. _{^{Contributed by: @ZiyueWang25, @XkunW}}

  uv run inspect eval inspect_evals/gdm_approved_directories
  uv run inspect eval inspect_evals/gdm_calculator_improvement
  uv run inspect eval inspect_evals/gdm_context_length_mod_instrumental_only
  uv run inspect eval inspect_evals/gdm_context_length_mod_irreversibility_only
  uv run inspect eval inspect_evals/gdm_database_tool
  uv run inspect eval inspect_evals/gdm_latency_calculator
  uv run inspect eval inspect_evals/gdm_max_messages_calculator
  uv run inspect eval inspect_evals/gdm_max_tokens
  uv run inspect eval inspect_evals/gdm_oversight_frequency
  uv run inspect eval inspect_evals/gdm_read_logs
  uv run inspect eval inspect_evals/gdm_turn_off_filters

• GDM Dangerous Capabilities: Stealth

  Test AI's ability to reason about and circumvent oversight. _{^{Contributed by: @ZiyueWang25}}

  uv run inspect eval inspect_evals/gdm_classifier_evasion
  uv run inspect eval inspect_evals/gdm_cover_your_tracks
  uv run inspect eval inspect_evals/gdm_oversight_pattern
  uv run inspect eval inspect_evals/gdm_strategic_rule_breaking

• DocVQA: A Dataset for VQA on Document Images

  DocVQA is a Visual Question Answering benchmark that consists of 50,000 questions covering 12,000+ document images. This implementation solves and scores the "validation" split. _{^{Contributed by: @evanmiller-anthropic}}

  uv run inspect eval inspect_evals/docvqa

• MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models

  A comprehensive dataset designed to evaluate Large Vision-Language Models (LVLMs) across a wide range of multi-image tasks. The dataset encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions. _{^{Contributed by: @Esther-Guo}}

  uv run inspect eval inspect_evals/mmiu

• V*Bench: A Visual QA Benchmark with Detailed High-resolution Images

  V*Bench is a visual question & answer benchmark that evaluates MLLMs in their ability to process high-resolution and visually crowded images to find and focus on small details. _{^{Contributed by: @bienehito}}

  uv run inspect eval inspect_evals/vstar_bench_attribute_recognition
  uv run inspect eval inspect_evals/vstar_bench_spatial_relationship_reasoning

• ZeroBench

  A lightweight visual reasoning benchmark that is (1) Challenging, (2) Lightweight, (3) Diverse, and (4) High-quality. _{^{Contributed by: @ItsTania}}

  uv run inspect eval inspect_evals/Zerobench
  uv run inspect eval inspect_evals/Zerobench Subquestions

• BBQ: Bias Benchmark for Question Answering

  A dataset for evaluating bias in question answering models across multiple social dimensions. _{^{Contributed by: @harshraj172, @shubhobm}}

  uv run inspect eval inspect_evals/bbq

• BOLD: Bias in Open-ended Language Generation Dataset

  A dataset to measure fairness in open-ended text generation, covering five domains: profession, gender, race, religious ideologies, and political ideologies. _{^{Contributed by: @harshraj172, @shubhobm}}

  uv run inspect eval inspect_evals/bold

• Personality

  An evaluation suite consisting of multiple personality tests that can be applied to LLMs. Its primary goals are twofold:

  1. Assess a model's default personality: the persona it naturally exhibits without specific prompting.
  2. Evaluate whether a model can embody a specified persona**: how effectively it adopts certain personality traits when prompted or guided. _{^{Contributed by: @guiem}}

  uv run inspect eval inspect_evals/personality_BFI
  uv run inspect eval inspect_evals/personality_TRAIT
  uv run inspect eval inspect_evals/personality_PRIME

• WritingBench: A Comprehensive Benchmark for Generative Writing

  A comprehensive evaluation benchmark designed to assess large language models' capabilities across diverse writing tasks. The benchmark evaluates models on various writing domains including academic papers, business documents, creative writing, and technical documentation, with multi-dimensional scoring based on domain-specific criteria. _{^{Contributed by: @jtv199}}

  uv run inspect eval inspect_evals/writingbench