datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
belebele
The Belebele Benchmark for Massively Multilingual NLU Evaluation
Belebele is a multiple-choice machine reading comprehension (MRC) dataset spanning 122 language variants. This dataset enables the evaluation of mono- and multi-lingual models in high-, medium-, and low-resource languages. Each question has four multiple-choice answers and is linked to a short passage from the FLORES-200 dataset. The human annotation procedure was carefully curated to create questions that discriminate… See the full description on the dataset page: https://huggingface.co/datasets/facebook/belebele.meta-active-readingS-EMBER
S-EMBER: A Large-Scale Benchmark for Streaming Egocentric Memory Retrieval
Episodic-memory video QA benchmark (face-blurred, audio-removed).
License & usage
This dataset is licensed under
CC BY-NC 4.0 and is provided
for non-commercial research use only. Access is gated: you must accept the
non-commercial terms above before downloading.
Contents
sember_mcq.jsonl — multiple-choice evaluation split.
sember_grounding.jsonl — answer-generation and… See the full description on the dataset page: https://huggingface.co/datasets/facebook/S-EMBER.wearable-ai
EgoWearBench Dataset (ECCV 2026)
Part of the Wearable AI Workshop at ECCV 2026.
A benchmark of egocentric (first-person, head-mounted wearable camera) videos paired with three complementary video question-answering tasks for evaluating wearable-AI assistants on real-world everyday activity videos.
▶ Baseline code & evaluation scripts: see starter_kit/README.md. The starter kit ships inside this repo, so git clone gives you the code and the data together.
Tasks… See the full description on the dataset page: https://huggingface.co/datasets/facebook/wearable-ai.natural_reasoningNaturalReasoning is a large-scale dataset for general reasoning tasks. It consists of high-quality challenging reasoning questions backtranslated from pretraining corpora DCLM and FineMath. The questions have been deduplicated and decontaminated from popular reasoning benchmarks including MATH, GPQA, MMLU-Pro, MMLU-STEM. For each question, we extract the reference final answer from the original document from the pretraining corpora if possible. We also provide a model-generated response from… See the full description on the dataset page: https://huggingface.co/datasets/facebook/natural_reasoning.cyberseceval3-visual-prompt-injection
Dataset Card for CyberSecEval 3 - Visual Prompt Injection Benchmark
Dataset Details
Dataset Description
This dataset provides a multimodal benchmark for visual prompt injection, with text/image inputs. It is part of CyberSecEval 3, the third edition of Meta's flagship suite of security benchmarks for LLMs to measure cybersecurity risks and capabilities across multiple domains.
Language(s): English
License: MIT
Dataset Sources
Repository: Link… See the full description on the dataset page: https://huggingface.co/datasets/facebook/cyberseceval3-visual-prompt-injection.recycling_the_web
Dataset Card for Recycling-The-Web Synthetic Data
We release 44.4B tokens of high-quality, model-filtered synthetic texts obtained via our REcycling the Web with guIded REwrite (REWIRE) approach.
The generation process involves taking all documents that are of moderate quality (i.e., having passed some rule-based filters),
using an LLM (Llama-3.3-70B-Instruct) to identify the purpose of the text content, and then asking the LLM to come up with an improved document conditioned on… See the full description on the dataset page: https://huggingface.co/datasets/facebook/recycling_the_web.multiloko
MultiLoKo: a multilingual local knowledge benchmark for LLMs
MultiLoKo is a multilingual knowledge benchmark, covering 30 languages plus English.
The questions are separately sourced for each language, with an annotation protocol designed to target locally relevant topics for the respective language.
MultiLoKo contains the original data for each language, as well as both human and machine-authored translations of each non-English subset into English and vice versa, facilitating… See the full description on the dataset page: https://huggingface.co/datasets/facebook/multiloko.BigOBench
👋 Overview
🚀 Introduction
📋 Getting Started with the data
🔥 problem_and_human_solutions_list.jsonl
🔥 complexity_labels_light.jsonl
🔥 complexity_labels_full.jsonl
🔥 time_complexity_test_set.jsonl
🔥 space_complexity_test_set.jsonl
License
📝 Citation
🚀 Introduction
BigO(Bench) is a benchmark of ~300 code problems to be solved in Python, along with 3,105 coding problems… See the full description on the dataset page: https://huggingface.co/datasets/facebook/BigOBench.facebook-xglm-564M-toksuite-detokenizedTraining data of the model detokenized in the exact order seen by the model.
The training data is partitioned into 8 chunks (chunk-0 through chunk-7), based on the GPU rank that generated the data. Each chunk contains detokenized text files in JSON Lines format (.jsonl).
principia-bench
Principia Bench
Principia Bench is a benchmark designed to evaluate language models' ability to derive mathematical objects from STEM-related problem statements. Each instance contains a problem statement and a ground-truth answer. The problem statements are drawn from four benchmarks—RealMath, SuperGPQA (with answers choices removed), Physics, and ARB—and are filtered to retain only instances where (1) the answer is expressed as a mathematical object, (2) the problem statement… See the full description on the dataset page: https://huggingface.co/datasets/facebook/principia-bench.face-to-face-Configoptimal_thinking_benchThis dataset is released as part of OptimalOptimalThinkingBench research project.
IMPORTANT: This is only a subset of OptimalThinkingBench that does not contain the math problems. To download the full dataset, please refer to our project materials here for more details.
Loading the dataset with transformers
This dataset is built using Llama-4-Maverick and Reasoning-Gym. Details on how to generate this dataset can be found in OptimalOptimalThinkingBench paper.
Minimal example below… See the full description on the dataset page: https://huggingface.co/datasets/facebook/optimal_thinking_bench.lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1645559101lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1646052073
GEM Submission
Submission name: Hugging Face test T5-base.outputs.json 36bf2a59
lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1646049601lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1645800191lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1646049378lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1646050898lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1646049876lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1645558682lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1646049424lewtun__hugging-face-test-t5-base.outputs.json-36bf2a59__1646051364llamafirewall-alignmentcheck-evals
Dataset Card for LlamaFirewall AlignmentCheck Evals
Dataset Details
Dataset Description
This dataset provides a dataset for prompt injection in an agentic environment. It is part of LlamaFirewall, an open-source security focused guardrail framework designed to serve as a final layer of defense against security risks associated with AI Agents. Specifically, this dataset is designed to evaluate the susceptibility of language models, and detect any misalignment… See the full description on the dataset page: https://huggingface.co/datasets/facebook/llamafirewall-alignmentcheck-evals.SACo-Gold
Dataset Card for SA-Co/Gold
SA-Co/Gold is a benchmark for promptable concept segmentation (PCS) in images. The benchmark contains images paired with text labels (also referred as Noun Phrases aka NPs), each annotated exhaustively with masks on all object instances that match the label. SA-Co/Gold comprises 7 subsets, each targeting a different annotation domain. For each subset, the annotations are multi-reviewed and agreed by 3 human annotators resulting in a high-quality… See the full description on the dataset page: https://huggingface.co/datasets/facebook/SACo-Gold.panda
Dataset Card for PANDA
Dataset Summary
PANDA (Perturbation Augmentation NLP DAtaset) consists of approximately 100K pairs of crowdsourced human-perturbed text snippets (original, perturbed). Annotators were given selected terms and target demographic attributes, and instructed to rewrite text snippets along three demographic axes: gender, race and age, while preserving semantic meaning. Text snippets were sourced from a range of text corpora (BookCorpus, Wikipedia, ANLI… See the full description on the dataset page: https://huggingface.co/datasets/facebook/panda.Wildchat-RIP-Filtered-by-8b-LlamaRIP is a method for perference data filtering. The core idea is that low-quality input prompts lead to high variance and low-quality responses. By measuring the quality of rejected responses and the reward gap between chosen and rejected preference pairs, RIP effectively filters prompts to enhance dataset quality.
We release 4k data that filtered from 20k Wildchat prompts. For each prompt, we provide 64 responses from Llama-3.1-8B-Instruct and their corresponding rewards obtained from ArmoRM.… See the full description on the dataset page: https://huggingface.co/datasets/facebook/Wildchat-RIP-Filtered-by-8b-Llama.SACo-VEval
SA-Co/VEval Dataset
License each domain has its own License
SA-Co/VEval - SA-V: CC-BY-NC 4.0
SA-Co/VEval - YT-Temporal-1B: CC-BY-NC 4.0
SA-Co/VEval - SmartGlasses: CC-by-4.0
SA-Co/VEval is an evaluation dataset comprising of 3 domains, each domain has a val and test split.
SA-Co/VEval - SA-V: videos are from the SA-V dataset
SA-Co/VEval - YT-Temporal-1B: videos are from the YT-Temporal-1B
SA-Co/VEval - SmartGlasses: egocentric videos from Smart Glasses
This Hugging Face dataset… See the full description on the dataset page: https://huggingface.co/datasets/facebook/SACo-VEval.Wildchat-RIP-Filtered-by-70b-LlamaRIP is a method for perference data filtering. The core idea is that low-quality input prompts lead to high variance and low-quality responses. By measuring the quality of rejected responses and the reward gap between chosen and rejected preference pairs, RIP effectively filters prompts to enhance dataset quality.
We release 4k data that filtered from 20k Wildchat prompts. For each prompt, we provide 32 responses from Llama-3.3-70B-Instruct and their corresponding rewards obtained from ArmoRM.… See the full description on the dataset page: https://huggingface.co/datasets/facebook/Wildchat-RIP-Filtered-by-70b-Llama.cs_facebook-comments
Dataset Card for Czech Facebook comments
Dataset Description
The dataset contains user comments from Facebook. Each comment contains text, sentiment (positive/negative/neutral).
The dataset has in total (train+validation+test) 6,600 reviews. The data is balanced.
Dataset Features
Each sample contains:
comment_id: unique string identifier of the comment.
sentiment_str: string representation of the rating - "pozitivní" / "neutrální" / "negativní"
sentiment_int:… See the full description on the dataset page: https://huggingface.co/datasets/fewshot-goes-multilingual/cs_facebook-comments.