Benchmark definition
HellaSwag
An adversarially filtered commonsense inference benchmark based on selecting plausible continuations to grounded situations.
Applicability does not mean evaluated; benchmark result/model score data is outside this UI contract.
About
- Description
- HellaSwag presents contexts from grounded activities/instructions with one plausible continuation and adversarially generated distractors designed to fool language models while remaining easy for humans.
- What it measures
- Grounded commonsense inference and plausible event continuation.
- Comparison guidance
- Not published
- Intended difficulty
- Not published
Not recorded for this benchmark: what it does not measure, why it matters, motivation, gap addressed.
Classification
- Categories
- Commonsense
- Domains
- Everyday / commonsense world
- Evaluation style
- Automatic reference matchStatic Q&A
Structure
One undivided item set. No components or scenarios recorded.
Tasks
None recorded
Metrics
What a result on this benchmark would be reported as. ctxdex does not carry results.
| Metric | Role | Direction | Unit |
|---|---|---|---|
Accuracy | Primary | Higher is better | Percent |
Methodology
Choose the real/plausible continuation from four alternatives; distractors are machine-generated and adversarially filtered.
Evaluation environment
What the benchmark's harness provides and allows. These are properties of the environment, not model requirements, and applicability does not test for them.
- Interaction
None. Static prompts and answers.
- Evaluation style
- Automatic reference match, Static Q&A
- Environment summary
- Not published
Model requirements
Required conditions decide applicability. Relevant ones describe what the benchmark exercises and decide nothing.
Required · decides applicability
Relevant · informational
CTXDEX models
30 of the 58 active, preview, experimental models meet this benchmark’s recorded requirements, at some scope. Applicable means the requirements are met, not that the model was evaluated.
View all 30 applicable models →- Applicable · full
- 30
- Applicable · components
- 0
- Applicable · scenarios
- 0
- Conditionally applicable
- 0
- Not recorded
- 28
Counted over 58 models in the default lifecycle scope (active, preview, experimental).
Every required condition, for the whole benchmark, is recorded as supported. Each line is one recorded condition and why it counts.
Every one of the 2 required conditions is recorded as supported.
MODALITY:INPUT:TEXTthe catalogue records SUPPORTED for TEXT (input)MODALITY:OUTPUT:TEXTthe catalogue records SUPPORTED for TEXT (output)
Every one of the 2 required conditions is recorded as supported.
MODALITY:INPUT:TEXTthe catalogue records SUPPORTED for TEXT (input)MODALITY:OUTPUT:TEXTthe catalogue records SUPPORTED for TEXT (output)
Applicability/relevance ordering only; never a performance ranking.
Scale
None recorded
Data and access
- Dataset
- Open
- Evaluation code
- Open source
- Unknown
- Leaderboard, Peer review, Maintenance
Coverage
- Languages, regions
None recorded
- Coverage summary
- Not published
Limitations
None recorded
Lineage
- Family
None recorded
- Related
None recorded
- Changes
None recorded
Sources
- HellaSwag paper ↗Research paper · Original creator
Freshness
- Content version
- 1.0
- Last reviewed
- 2026-09-22
- Refresh model
- Static
- Time sensitivity
- Low
- Benchmark data
- c40cccac878af5d0
13 contract fields are empty for every benchmark in this release and are not shown: applicability groups, conditional requirements, feature requirements, model relevance domains, required capabilities, openness summary, reference artifacts, interaction protocols, tool protocols, benchmark last major refresh at, short name, target expertise level, target population.