HellaSwag
DatasetFree70K commonsense reasoning questions with adversarial distractors.
Capabilities5 decomposed
adversarial-filtered multiple-choice evaluation
Medium confidenceEvaluates model reasoning by presenting 70,000 multiple-choice questions where incorrect options were generated by language models and adversarially selected to fool machines while remaining obviously wrong to humans. The filtering process identifies plausible-but-incorrect continuations that expose gaps in commonsense reasoning, creating a harder benchmark than human-authored distractors. Models must select the single correct continuation from four options, with evaluation metrics tracking accuracy against human baseline (95.6%).
Uses adversarial filtering where incorrect options are generated by language models and selected specifically because they fool machines while remaining obvious to humans, rather than relying on human-authored distractors. This creates a harder, more realistic benchmark that exposes model weaknesses in distinguishing plausible-but-wrong continuations.
Harder and more realistic than manually-authored multiple-choice benchmarks (e.g., RACE, SWAG) because adversarial distractors target actual model failure modes rather than generic wrong answers, making it a better predictor of real-world commonsense reasoning gaps.
physical-commonsense continuation prediction
Medium confidenceEvaluates models' ability to predict the most plausible next action or outcome in everyday physical scenarios (e.g., 'person is hammering a nail, what happens next?'). The dataset includes video-grounded scenarios where the correct continuation is the actual next frame or action from real video, and the model must choose among four options. This tests understanding of physics, object interactions, and temporal causality in real-world activities.
Grounds scenarios in real video sequences where the correct answer is the actual next frame/action from the video, rather than synthetic or hypothetical continuations. This ensures ground truth is tied to real-world physics and human behavior, not annotator preferences.
More grounded in real-world physics than synthetic commonsense benchmarks (e.g., ATOMIC, ConceptNet) because correct answers are actual video continuations, making it a stronger test of whether models truly understand physical causality vs. memorizing common-sense patterns.
social-understanding and temporal-reasoning evaluation
Medium confidenceAssesses models' ability to understand social interactions, emotional context, and temporal sequences in everyday scenarios. The dataset includes questions about social dynamics (e.g., 'person is arguing with friend, what happens next?') and temporal reasoning (e.g., 'person is putting on shoes, what's the next step?'). Models must select the most plausible continuation from four options, testing understanding of social norms, emotional progression, and action sequences.
Combines social dynamics and temporal reasoning in a single benchmark, with scenarios grounded in real video where social interactions and action sequences are captured. Adversarial filtering specifically targets model weaknesses in understanding social norms and temporal causality.
Covers both social and temporal reasoning in one dataset, whereas most commonsense benchmarks (e.g., CommonsenseQA, CSQA) focus primarily on static knowledge; the video grounding ensures social scenarios reflect real human behavior rather than annotator assumptions.
model-capability benchmarking against human baseline
Medium confidenceProvides a standardized evaluation framework comparing model performance against a human baseline (95.6% accuracy) on commonsense reasoning tasks. The dataset includes 70,000 examples with ground truth labels, enabling researchers to track whether their models are approaching or exceeding human-level performance. Evaluation is straightforward: compute accuracy on the full dataset or subsets, then compare against the human baseline and other published models.
Provides a human baseline (95.6%) derived from actual human annotators, enabling researchers to measure progress toward human-level performance. The adversarial filtering ensures the benchmark remains challenging even as frontier models improve, preventing ceiling effects.
More challenging and realistic than generic multiple-choice benchmarks because adversarial filtering targets actual model weaknesses; human baseline is well-established and published, making it easier to contextualize model performance than on benchmarks with unknown or variable human performance.
adversarial-robustness evaluation through machine-generated distractors
Medium confidenceTests model robustness by using language-model-generated incorrect options that are specifically selected to fool machines. Rather than relying on human-authored distractors (which may be obviously wrong), the dataset uses adversarial filtering to identify machine-generated options that are plausible to models but clearly wrong to humans. This reveals whether models are truly reasoning or merely pattern-matching, and identifies specific failure modes where models confuse plausible-but-incorrect continuations with correct ones.
Uses adversarial filtering to select machine-generated distractors that fool models while remaining obviously wrong to humans, rather than using generic or human-authored wrong answers. This creates a benchmark that specifically targets model weaknesses in distinguishing plausible-but-incorrect continuations.
More effective at revealing model reasoning shortcuts than benchmarks with human-authored distractors, because adversarial filtering identifies exactly which types of plausible-but-wrong answers fool machines, enabling targeted robustness evaluation and improvement.
Capabilities are decomposed by AI analysis. Each maps to specific user intents and improves with match feedback.
Related Artifactssharing capabilities
Artifacts that share capabilities with HellaSwag, ranked by overlap. Discovered automatically through the match graph.
RealWorldQA
Real-world visual QA requiring spatial reasoning.
LMSYS Chatbot Arena
Crowdsourced LLM evaluation — side-by-side blind voting, Elo ratings, most trusted LLM benchmark.
Tree of Thoughts: Deliberate Problem Solving with Large Language Models (ToT)
* ⭐ 05/2023: [LIMA: Less Is More for Alignment (LIMA)](https://arxiv.org/abs/2305.11206)
WinoGrande
44K pronoun resolution problems testing commonsense understanding.
HotpotQA
113K questions requiring multi-hop reasoning across Wikipedia articles.
TriviaQA
95K trivia questions requiring cross-document reasoning.
Best For
- ✓LLM researchers evaluating model capabilities on commonsense reasoning
- ✓teams building embodied AI or robotics systems requiring physical understanding
- ✓practitioners assessing whether fine-tuning or architectural changes improve reasoning
- ✓benchmark suite maintainers needing a standardized, adversarially-robust evaluation
- ✓robotics teams building systems that must predict physical outcomes of actions
- ✓video understanding researchers evaluating temporal reasoning
- ✓embodied AI developers testing commonsense physics understanding
- ✓multimodal model researchers assessing vision-language reasoning on physical tasks
Known Limitations
- ⚠Adversarial filtering may introduce subtle biases toward specific model architectures or training approaches
- ⚠70,000 examples may not cover all commonsense domains (e.g., specialized technical or cultural knowledge)
- ⚠Multiple-choice format constrains evaluation to selection rather than open-ended generation or explanation
- ⚠Human baseline (95.6%) is not 100%, indicating some ambiguity in ground truth labels
- ⚠No fine-grained error analysis built-in — requires external tooling to categorize failure modes
- ⚠Video grounding is implicit (via scenario text) rather than explicit visual input — models cannot use pixel-level features
Requirements
Input / Output
UnfragileRank
UnfragileRank is computed from adoption signals, documentation quality, ecosystem connectivity, match graph feedback, and freshness. No artifact can pay for a higher rank.
About
Commonsense reasoning benchmark with 70,000 multiple-choice questions requiring models to select the most plausible continuation of everyday scenarios. Uses adversarial filtering: incorrect options were generated by language models and selected specifically because they fool machines while being obvious to humans. Tests physical commonsense (what happens next in activities), social understanding, and temporal reasoning. Human accuracy is 95.6%; frontier LLMs now approach this level.
Categories
Alternatives to HellaSwag
The GitHub for AI — 500K+ models, datasets, Spaces, Inference API, hub for open-source AI.
Compare →FLUX, Stable Diffusion, SDXL, SD3, LoRA, Fine Tuning, DreamBooth, Training, Automatic1111, Forge WebUI, SwarmUI, DeepFake, TTS, Animation, Text To Video, Tutorials, Guides, Lectures, Courses, ComfyUI, Google Colab, RunPod, Kaggle, NoteBooks, ControlNet, TTS, Voice Cloning, AI, AI News, ML, ML News,
Compare →Are you the builder of HellaSwag?
Claim this artifact to get a verified badge, access match analytics, see which intents users search for, and manage your listing.
Get the weekly brief
New tools, rising stars, and what's actually worth your time. No spam.
Data Sources
Looking for something else?
Search →