HumanEval: 164 Programming Tasks
A benchmark of 164 Python tasks with signatures, docstrings, and unit tests for code generation evaluation.
- English
Hugging Face datasets with plain-language descriptions: what's inside and what models to train on them. 4,788 in the catalog so far, with new ones added every hour.
Answered by Gemini 4 · the chat is kept only in this tab
A benchmark of 164 Python tasks with signatures, docstrings, and unit tests for code generation evaluation.
24.3 million interleaved image-text examples in English and Chinese for training multimodal models.
About 1,000 Python tasks with descriptions, solutions, and tests for code generation evaluation.
Multiple-choice sentence completion tasks to evaluate commonsense reasoning in language models.
English and German text scripts with sudden emotional shifts for a single character to train expressive TTS models.
Large-scale English corpus of 1.61 million web novel records containing text, tables, and images.
Environment suite for training LLM agents with checks in code, cybersecurity, web, and music.
Mirror of Terminal-Bench 2.1 benchmark tasks adapted for execution in Harbor.
Over 10,000 hours of real-world recordings of craftsmanship, cooking, and construction work with IMU sensor data.
Over 1.1 million hours of web audio in 100+ languages for ASR, TTS, and audio representation learning.
A test split from the well-known ObjectNet benchmark for evaluating computer vision models without background and viewpoint bias.
4 trillion tokens and 3 billion documents for LLM pre-training; strong benchmark results.
A massive dataset of over 1.1 million hours of audio across 100+ languages for speech tasks.
Speech data in 102 languages for training and evaluating ASR on parallel sentences.
A subset of 500 math problems from the MATH benchmark for evaluating step-by-step reasoning.
Synthetic short stories with a limited English vocabulary for training and evaluating language models.
Synthesized and processed articles with metadata in 5 languages for text generation, translation, and summarization tasks.
Over 10 thousand Arabic books in PDF, TXT, and DOCX formats, recognized via Google Document AI.
English MINT-1T HTML corpus: part of an open 1 trillion token and 3.4 billion image multimodal dataset.
64 thousand real image-question-answer pairs from StackExchange for VQA tasks with verified answers.
FineWeb-2 pretraining corpus with 1000+ languages for training language models.
A compact, versioned multimodal benchmark for testing models on terminal tasks.
480 tasks from bankers, consultants, and lawyers to test AI agents in cross-application scenarios.
Over 215 hours of recordings from 41 cameras and tactile grasps interacting with 198 deformable objects.