Examples
The examples below demonstrate a variety of evaluation types and techniques. If you have just begun learning Inspect, you might benefit from reviewing the Tutorial examples before exploring these.
Coding
-
Evaluating correctness for synthesizing Python programs from docstrings. Demonstrates custom scorers and sandboxing untrusted model code.
-
Measuring the ability of these models to synthesize short Python programs from natural language descriptions. Demonstrates custom scorers and sandboxing untrusted model code.
-
Software engineering problems drawn from real GitHub issues and corresponding pull requests across 12 popular Python repositories. Demonstrates sandboxing untrusted model code.
Cybersecurity
-
Measure expertise in coding, cryptography (i.e. binary exploitation, forensics), reverse engineering, and recognizing security vulnerabilities. Demonstrates tool use and sandboxing untrusted model code.
-
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.
Mathematics
-
Dataset of 12,500 challenging competition mathematics problems. Demonstrates fewshot prompting and custom scorers.
-
Dataset of 8.5K high quality linguistically diverse grade school math word problems. Demostrates fewshot prompting.
-
Diverse mathematical and visual tasks that require fine-grained, deep visual understanding and compositional reasoning. Demonstrates multimodal inputs and custom scorers.
Reasoning
-
Dataset of natural, grade-school science multiple-choice questions (authored for human tests).
-
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.
-
Measure physical commonsense reasoning (e.g. "To apply eyeshadow without a brush, should I use a cotton swab or a toothpick?")
-
Reading comprehension dataset that queries for complex, non-factoid information, and require difficult entailment-like inference to solve.
-
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).
-
Set of 273 expert-crafted pronoun resolution problems originally designed to be unsolvable for statistical models that rely on selectional preferences or word associations.
-
Reading comprehension tasks collected from the English exams for middle and high school Chinese students in the age range between 12 to 18.
-
Multimodal questions from college exams, quizzes, and textbooks, covering six core disciplinestasks, demanding college-level subject knowledge and deliberate reasoning. Demonstrates multimodel inputs.
-
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.
-
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.
-
Questions from human-centric standardized exams, such as college entrance exams, law school admission tests, math competitions, and lawyer qualification tests. Demonstrates custom scoring.
Knowledge
-
Evaluate models on 57 tasks including elementary mathematics, US history, computer science, law, and more.
-
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.
-
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).
-
Measure question answering with commonsense prior knowledge.
-
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.
-
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.
-
Novel biomedical question answering (QA) dataset collected from PubMed abstracts.
No matching items