HellaSwag

HellaSwag
Overview
Full name	Harder Endings, Longer contexts, and Low-shot Activities for Situations With Adversarial Generations
Abbreviation	HellaSwag
Description	A challenging commonsense reasoning benchmark using adversarial filtering to test physical understanding in language models
Release date	2019-05-19
Latest version	1.0
Benchmark updated	2019
Authors	Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, Yejin Choi
Organization	Allen Institute for AI, University of Washington, Carnegie Mellon University
Technical Details
Type	Commonsense Reasoning, Natural Language Inference, Sentence Completion
Modality	Text
Task format	Multiple-choice sentence completion
Number of tasks	59,950
Total examples	59,950 questions (39,905 train, 10,042 val, 10,003 test)
Evaluation metric	Accuracy
Domains	ActivityNet, WikiHow
Languages	English
Performance
Human performance	95.6%
Baseline	<48% (BERT-Large, 2019) Property "Baseline score" (as page type) with input value "" contains invalid characters or is incomplete and therefore can cause unexpected results during a query or annotation process.
SOTA score	95.3%
SOTA model	GPT-4 (10-shot)
SOTA date	2024
Saturated	Nearly (GPT-4 matches human performance)
Resources
Website	Official website
Paper	Paper
GitHub	Repository
Dataset	Download
License	MIT
Predecessor	SWAG
Successor	HellaSwag-Pro

HellaSwag (Harder Endings, Longer contexts, and Low-shot Activities for Situations With Adversarial Generations) is a challenging commonsense reasoning benchmark designed to evaluate language models' understanding of physical situations and everyday activities. Created by researchers from the Allen Institute for AI, University of Washington, and Carnegie Mellon University and published at ACL 2019^[1], HellaSwag uses a novel adversarial filtering technique to create sentence completion tasks that are trivial for humans (95.6% accuracy) but initially proved extremely challenging for state-of-the-art models (<48% accuracy). The benchmark has become a standard evaluation for measuring progress in physical commonsense understanding, with modern models like GPT-4 finally achieving human-level performance.

Overview

HellaSwag represents a significant advancement in evaluating true commonsense understanding in AI systems. Unlike benchmarks that can be solved through pattern matching or memorization, HellaSwag specifically targets the "Goldilocks zone" of complexity, generating wrong answers that are superficially plausible and contain expected words but violate physical commonsense in ways obvious to humans^[2]. This approach revealed that models achieving impressive performance on other benchmarks fundamentally lacked understanding of basic physical interactions and causal relationships.

The Commonsense Challenge

HellaSwag addresses a critical question in AI: Can machines truly understand and reason about the physical world, or do they merely memorize statistical patterns? The benchmark focuses on:

• **Physical causality**: Understanding cause-and-effect in real-world scenarios
• **Temporal reasoning**: Predicting what happens next in sequences of events
• **Activity understanding**: Comprehending how everyday tasks unfold
• **Contextual coherence**: Maintaining logical consistency across sentences

Adversarial Filtering Methodology

The Creation Process

HellaSwag's key innovation lies in its sophisticated adversarial filtering (AF) technique^[1]:

Stage	Process	Purpose
**1. Generation**	GPT generates candidate wrong answers	Create plausible-sounding distractors
**2. Discrimination**	BERT-Large filters easy-to-detect options	Remove obvious wrong answers
**3. Iteration**	Multiple rounds of generation and filtering	Increase difficulty progressively
**4. Human Validation**	Crowd workers evaluate final options	Ensure human solvability
**5. Selection**	Top 59,950 most challenging examples retained	Create final dataset

The Goldilocks Zone

The adversarial filtering targets a specific difficulty profile:

Characteristic	Human Perception	Model Perception
**Grammaticality**	Correct	Correct
**Vocabulary**	Appropriate	Appropriate
**Surface plausibility**	Somewhat believable	Very believable
**Physical coherence**	Obviously wrong	Unclear
**Causal logic**	Violated	Not detected

Dataset Structure

Composition and Sources

HellaSwag draws from two primary sources to ensure diversity:

Source	Examples	Domain	Difficulty for Models
**ActivityNet Captions**	~25,000	Video descriptions	Moderate
**WikiHow**	~45,000	How-to articles	Very Hard
**Total**	59,950	Mixed	Challenging

Data Splits

Split	Size	Purpose	Evaluation Type
**Training**	39,905	Model training	Standard
**Validation**	10,042	Hyperparameter tuning	In-domain & out-of-domain
**Test**	10,003	Final evaluation	In-domain & out-of-domain

Task Format

Each HellaSwag example consists of:

``` {

 "activity_label": "Making a sandwich",
 "context": "A person takes two slices of bread from a loaf. They open a jar of peanut butter.",
 "endings": [
   "They spread peanut butter on one slice.",  // Correct
   "They put the jar in the refrigerator and walk away.",  // Adversarial
   "They start cutting the bread into small pieces.",  // Adversarial
   "They throw the bread in the trash."  // Adversarial
 ],
 "label": 0

} ```

Performance Evolution

Historical Performance (2019)

At release, HellaSwag revealed a massive performance gap:

Model	Accuracy	Gap to Human
Humans	95.6%	-
BERT-Large	47.3%	48.3%
GPT (original)	41.7%	53.9%
LSTM + ELMo	38.0%	57.6%
Random Baseline	25.0%	70.6%

Current Performance (2024-2025)

Modern models have largely closed the gap:

Model	Setting	Accuracy	Year
GPT-4	10-shot	95.3%	2024
Humans	-	95.6%	-
Claude-3 Opus	Few-shot	~93%	2024
GPT-4o	Few-shot	88-90%	2024
GPT-3.5	10-shot	85.5%	2023
Falcon-40B	Fine-tuned	85.3%	2023
LLaMA-2-70B	Fine-tuned	~82%	2023

Comparison with SWAG

Evolution from SWAG to HellaSwag

HellaSwag builds upon and improves the earlier SWAG benchmark:

Aspect	SWAG	HellaSwag
**Size**	113,000 examples	70,000 examples
**Generation Model**	Simple LM	GPT
**Filtering Models**	Basic classifiers	BERT-Large ensemble
**Human Accuracy**	88%	95.6%
**Initial SOTA**	~86%	<48%
**Difficulty**	Moderate	High
**Focus**	General situations	Physical commonsense

Cross-Training Analysis

Experiments revealed the increased difficulty of HellaSwag<sup>[1]</sup>:

Training	Evaluation	Performance Drop
SWAG	HellaSwag	-12%
HellaSwag	SWAG	-15%
Joint training	Both	Best overall

Key Findings and Insights

Domain-Specific Analysis

Performance varies significantly across domains:

Domain	Human Accuracy	BERT-Large (2019)	Difficulty Factor
ActivityNet	94.1%	53.3%	Moderate
WikiHow	96.5%	45.0%	Very High
Out-of-domain	95.2%	35.6%	Extreme

Error Analysis

Common failure modes in early models included:

1. **Temporal confusion**: Mixing up order of events 2. **Physical impossibilities**: Suggesting actions that violate physics 3. **Context ignorance**: Completions that ignore established context 4. **Causal breaks**: Missing cause-effect relationships 5. **Object permanence**: Forgetting about mentioned objects

Impact on NLP Research

Methodological Contributions

HellaSwag introduced several influential concepts:

Contribution	Description	Adoption
**Adversarial Filtering**	Using models to create hard examples	Widely adopted
**Human-in-the-loop validation**	Ensuring human solvability	Standard practice
**Goldilocks complexity**	Targeting optimal difficulty	Design principle
**Physical commonsense focus**	Emphasizing real-world understanding	Research direction

Spawned Research

HellaSwag has inspired numerous follow-up works:

• **HellaSwag-Pro (2024)**: Bilingual extension with 11,200 cases
• **GoldenSwag**: Corrected subset addressing validity issues
• **Physical reasoning benchmarks**: PIGLeT, PIQA, and others
• **Adversarial dataset creation**: Methodology applied across domains

Limitations and Criticisms

Identified Issues

Recent research has identified several limitations<sup>[3]</sup>:

Issue	Description	Impact
**Grammatical errors**	Some examples contain typos	Affects 5-10% of data
**Multiple valid answers**	Some questions have >1 correct option	Evaluation ambiguity
**Context ambiguity**	Unclear or incomplete contexts	Interpretation variance
**Construct validity**	May not fully capture commonsense	Theoretical concern

Saturation Concerns

With GPT-4 achieving near-human performance, questions arise about:

1. **Benchmark saturation**: Is HellaSwag still useful for evaluation? 2. **True understanding**: Do high scores indicate genuine commonsense? 3. **Generalization**: Performance on real-world tasks vs. benchmark 4. **Need for evolution**: Requirements for next-generation benchmarks

Technical Implementation

Using HellaSwag

```python from datasets import load_dataset

1. Load HellaSwag dataset

dataset = load_dataset("Rowan/hellaswag")

1. Access splits

train_data = dataset['train'] val_data = dataset['validation'] test_data = dataset['test']

1. Example evaluation

def evaluate_model(model, data):

   correct = 0
   for example in data:
       context = example['ctx']
       endings = example['endings']
       label = example['label']
       
       # Model predicts best ending
       prediction = model.predict(context, endings)
       
       if prediction == label:
           correct += 1
   
   accuracy = correct / len(data)
   return accuracy

```

Evaluation Protocols

Protocol	Description	Use Case
**Zero-shot**	No training examples	Test true generalization
**Few-shot**	1-10 examples	Standard evaluation
**Fine-tuned**	Full training set	Maximum performance
**Out-of-domain**	Held-out categories	Generalization testing

Future Directions

Evolution and Extensions

Direction	Description	Status
**Multilingual versions**	Extending beyond English	In progress (HellaSwag-Pro)
**Harder adversarial filtering**	More sophisticated generation	Research ongoing
**Multimodal integration**	Adding visual context	Proposed
**Dynamic updating**	Continuous difficulty adjustment	Conceptual

Significance

HellaSwag has played a crucial role in advancing our understanding of commonsense reasoning in AI systems. By revealing the gap between surface-level language proficiency and genuine physical understanding, it pushed the field toward developing models with deeper comprehension capabilities. The benchmark's adversarial filtering methodology has become a standard approach for creating challenging evaluations, influencing dataset design across NLP.

While modern models have largely conquered HellaSwag, achieving human-level performance, the benchmark's legacy continues through its methodological contributions and the research directions it inspired. The journey from <48% to >95% accuracy represents not just numerical progress but fundamental improvements in how language models understand and reason about the physical world, a capability essential for AI systems that interact with real-world environments.

See Also

• SWAG
• Commonsense Reasoning
• Adversarial Filtering
• Allen Institute for AI
• Physical Commonsense
• Natural Language Inference
• Benchmark Saturation
• GPT-4

References

1. Zellers, R., Holtzman, A., Bisk, Y., Farhadi, A., & Choi, Y. (2019). "HellaSwag: Can a Machine Really Finish Your Sentence?" Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL 2019). https://arxiv.org/abs/1905.07830
2. Zellers, R., Bisk, Y., Schwartz, R., & Choi, Y. (2018). "SWAG: A Large-Scale Adversarial Dataset for Grounded Commonsense Inference." Proceedings of EMNLP 2018. https://arxiv.org/abs/1808.05326
3. Caba Heilbron, F., Escorcia, V., Ghanem, B., & Carlos Niebles, J. (2015). "ActivityNet: A Large-Scale Video Benchmark for Human Activity Understanding." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
4. Brown, T. B., Mann, B., Ryder, N., et al. (2020). "Language Models are Few-Shot Learners." Proceedings of NeurIPS 2020. https://arxiv.org/abs/2005.14165
5. Gao, L., Tow, J., Abbasi, B., et al. (2024). "A Framework for Few-Shot Language Model Evaluation." EleutherAI. https://github.com/EleutherAI/lm-evaluation-harness
6. Hugging Face. (2024). "Open LLM Leaderboard v2." https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard
7. Surge AI. (2022). "HellaSwag or HellaBad? 36% of this popular LLM benchmark contains errors." https://surgehq.ai/blog/hellaswag-or-hellabad-36-of-this-popular-llm-benchmark-contains-errors
8. Chizhov, P., Nee, M., Langlais, P.-C., & Yamshchikov, I. P. (2025). "What the HellaSwag? On the Validity of Common-Sense Reasoning Benchmarks." https://arxiv.org/abs/2504.07825
9. Sainz, O., Campos, J., Garcia-Ferrero, I., et al. (2024). "How Much Can We Forget About Data Contamination?" https://arxiv.org/abs/2410.03249
10. Xu, Z., et al. (2025). "How Contaminated Is Your Benchmark? Quantifying Dataset Leakage in Large Language Models with Kernel Divergence." https://arxiv.org/abs/2502.00678
11. Wang, Q., et al. (2025). "HellaSwag-Pro: A Large-Scale Bilingual Benchmark for Evaluating the Robustness of LLMs in Commonsense Reasoning." Findings of ACL 2025. https://arxiv.org/abs/2502.11393
12. Bisk, Y., Zellers, R., Le Bras, R., Gao, J., & Choi, Y. (2020). "PIQA: Reasoning about Physical Intuition in Natural Language." Proceedings of AAAI 2020.
13. Sakaguchi, K., Le Bras, R., Bhagavatula, C., & Choi, Y. (2020). "WinoGrande: An Adversarial Winograd Schema Challenge at Scale." Proceedings of AAAI 2020.

Cite error: <ref> tag with name "hellaswag_github" defined in <references> is not used in prior text.
Cite error: <ref> tag with name "hellaswag_leaderboard" defined in <references> is not used in prior text.