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

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

First, install the inspect_ai and inspect_evals Python packages with:

pip install inspect_ai
pip install git+https://github.com/UKGovernmentBEIS/inspect_evals

If developing locally, you can install the packages in editable mode with:

pip install -e ".[dev]"

You will also need to install any packages required to interact with the models you'd like to evaluate and define any required API key environment variables. For example, to work with OpenAI models:

export OPENAI_API_KEY=<openai-api-key>
pip install openai

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

   pip install "inspect_evals[swe_bench] @ git+https://github.com/UKGovernmentBEIS/inspect_evals"
   pip install -e ".[swe_bench]" # If developing on the pacakge locally

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

inspect eval inspect_evals/gpqa --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 direcotry 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, AzureAI, AWS Bedrock, TogetherAI, Groq, HuggingFace, vLLM, Ollama, and more. See the Model Providers documentation for additional details.

• HumanEval: Evaluating Large Language Models Trained on Code

  Evaluating correctness for synthesizing Python programs from docstrings. Demonstrates custom scorers and sandboxing untrusted model code. _{^{Contributed by: @adil-a}}

  inspect eval inspect_evals/humaneval
  

• MBPP: Mostly Basic Python Problems

  Measuring the ability of these models to synthesize short Python programs from natural language descriptions. Demonstrates custom scorers and sandboxing untrusted model code. _{^{Contributed by: @jddantes}}

  inspect eval inspect_evals/mbpp
  

• SWE-Bench: Resolving Real-World GitHub Issues

  Software engineering problems drawn from real GitHub issues and corresponding pull requests across 12 popular Python repositories. Demonstrates sandboxing untrusted model code. _{^{Contributed by: @max-kaufmann}}

  inspect eval inspect_evals/swe_bench
  

• GAIA: A Benchmark for General AI Assistants

  GAIA 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}}

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

• InterCode: Capture the Flag

  Measure expertise in coding, cryptography (i.e. binary exploitation, forensics), reverse engineering, and recognizing security vulnerabilities. Demonstrates tool use and sandboxing untrusted model code. _{^{Contributed by: @jjallaire}}

  inspect eval inspect_evals/gdm_intercode_ctf
  

• 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}}

  inspect eval inspect_evals/gdm_in_house_ctf
  

• AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents

  Diverse set of 110 explicitly malicious agent tasks (440 with augmentations), covering 11 harm categories including fraud, cybercrime, and harassment. _{^{Contributed by: @alexandrasouly-aisi, @EricWinsorDSIT, @max-andr, @xanderdavies}}

  inspect eval inspect_evals/agentharm
  inspect eval inspect_evals/agentharm_benign
  

• 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}}

  inspect eval inspect_evals/wmdp_bio
  inspect eval inspect_evals/wmdp_chem
  inspect eval inspect_evals/wmdp_cyber
  

• MATH: Measuring Mathematical Problem Solving

  Dataset of 12,500 challenging competition mathematics problems. Demonstrates fewshot prompting and custom scorers. _{^{Contributed by: @xeon27}}

  inspect eval inspect_evals/math
  

• GSM8K: Training Verifiers to Solve Math Word Problems

  Dataset of 8.5K high quality linguistically diverse grade school math word problems. Demostrates fewshot prompting. _{^{Contributed by: @jjallaire}}

  inspect eval inspect_evals/gsm8k
  

• MathVista: Evaluating Mathematical Reasoning in Visual Contexts

  Diverse mathematical and visual tasks that require fine-grained, deep visual understanding and compositional reasoning. Demonstrates multimodal inputs and custom scorers. _{^{Contributed by: @ShivMunagala}}

  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}}

  inspect eval inspect_evals/arc_easy
  inspect eval inspect_evals/arc_challenge
  

• HellaSwag: Can a Machine Really Finish Your Sentence?

  Evaluting commonsense natural language inference: given an event description such as "A woman sits at a piano," a machine must select the most likely followup. _{^{Contributed by: @jjallaire}}

  inspect eval inspect_evals/hellaswag
  

• PIQA: Reasoning about Physical Commonsense in Natural Language

  Measure physical commonsense reasoning (e.g. "To apply eyeshadow without a brush, should I use a cotton swab or a toothpick?") _{^{Contributed by: @seddy-aisi}}

  inspect eval inspect_evals/piqa
  

• 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}}

  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}}

  inspect eval inspect_evals/drop
  

• 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}}

  inspect eval inspect_evals/winogrande
  

• 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}}

  inspect eval inspect_evals/race_h
  

• MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark

  Multimodal questions from college exams, quizzes, and textbooks, covering six core disciplinestasks, demanding college-level subject knowledge and deliberate reasoning. Demonstrates multimodel inputs. _{^{Contributed by: @shaheenahmedc}}

  inspect eval inspect_evals/mmmu_multiple_choice
  inspect eval inspect_evals/mmmu_open
  

• 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}}

  inspect eval inspect_evals/squad
  

• IFEval: Instruction-Following Evaluation for Large Language Models

  Evaluates the ability to follow a set of "verifiable instructions" such as "write in more than 400 words" and "mention the keyword of AI at least 3 times. Demonstrates custom scoring. _{^{Contributed by: @adil-a}}

  inspect eval inspect_evals/ifeval
  

• MMLU: Measuring Massive Multitask Language Understanding

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

  inspect eval inspect_evals/mmlu
  

• MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark

  An enhanced dataset designed to extend the mostly knowledge-driven MMLU benchmark by integrating more challenging, reasoning-focused questions and expanding the choice set from four to ten options. _{^{Contributed by: @xeon27}}

  inspect eval inspect_evals/mmlu_pro
  

• GPQA: A Graduate-Level Google-Proof Q&A Benchmark

  Challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry (experts at PhD level in the corresponding domains reach 65% accuracy). _{^{Contributed by: @jjallaire}}

  inspect eval inspect_evals/gpqa_diamond
  

• CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge

  Measure question answering with commonsense prior knowledge. _{^{Contributed by: @jjallaire}}

  inspect eval inspect_evals/commonsense_qa
  

• 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}}

  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}}

  inspect eval inspect_evals/xstest
  

• PubMedQA: A Dataset for Biomedical Research Question Answering

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

  inspect eval inspect_evals/pubmedqa
  

• 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.

  inspect eval inspect_evals/agie_aqua_rat
  inspect eval inspect_evals/agie_logiqa_en
  inspect eval inspect_evals/agie_lsat_ar
  inspect eval inspect_evals/agie_lsat_lr