WinoGrande
Binary fill-in-the-blank commonsense problems, scaled from Winograd schemas and adversarially filtered to reduce dataset shortcuts.
The original XL result was 79.1% against a 94.0% human estimate, leaving real headroom, but the static public inputs and widespread validation-set use make modern scores contamination-sensitive.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| Llama 3.1 405B (5-shot) | 89.2% | 2024-07-23 | vendor-reported | Source ↗ |
| Llama 3 70B (5-shot) | 83.1% | 2024-04-18 | vendor-reported | Source ↗ |
| Claude 3 Opus (5-shot) | 88.5% | 2024-03-04 | vendor-reported | Source ↗ |
| Llama 2 70B (5-shot) | 81.8% | 2023-07-18 | vendor-reported | Source ↗ |
| GPT-4 (5-shot) | 87.5% | 2023-03-14 | vendor-reported | Source ↗ |
| PaLM 540B (5-shot) | 85.1% | 2022-04-04 | vendor-reported | Source ↗ |
| GPT-3 175B (few-shot) | 77.7% | 2020-07-22 | vendor-reported | Source ↗ |
| BERT-Large (fine-tuned) | 64.9% | 2019-07-24 | independent | Source ↗ |
| RoBERTa-large (fine-tuned) | 79.1% | 2019-07-24 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Human accuracy on pronoun coreference resolution and adversarial commonsense pairs.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.accuracy (%)Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.$5 – $20 USD for full benchmark evaluation run on frontier APIs.
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.lm_eval --model hf --model_args pretrained=<model_path> --tasks winogrande --batch_size autoopencompass --datasets winogrande --models <model_config># Standard API Evaluation Loop
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}],
temperature=0.0,
)When publishing results for WinoGrande, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.
Contamination & Memorization Analysis
Audit of pretraining exposure risks, memorization vectors, and refresh policies.Static fixed snapshot
Public on web / HuggingFace