SCROLLS
Seven-task long-sequence suite covering summarization, QA, and NLI over naturally long English texts.
SCROLLS is still a standard long-sequence NLP suite with an official submission flow, but for frontier LLMs it is partly superseded by ZeroSCROLLS and newer very-long-context benchmarks.
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 |
|---|---|---|---|---|
| Naive (paper Table 2 Avg) | 19.35% | 2022-01-10 | independent | Source ↗ |
| BART 256 (paper Table 2 Avg) | 26.35% | 2022-01-10 | independent | Source ↗ |
| BART 512 (paper Table 2 Avg) | 27.58% | 2022-01-10 | independent | Source ↗ |
| BART 1024 (paper Table 2 Avg) | 29.01% | 2022-01-10 | independent | Source ↗ |
| LED 1024 (paper Table 2 Avg) | 27.06% | 2022-01-10 | independent | Source ↗ |
| LED 4096 (paper Table 2 Avg) | 28.3% | 2022-01-10 | independent | Source ↗ |
| LED 16384 (paper Table 2 Avg) | 29.16% | 2022-01-10 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Composite human performance across 7 long-document NLP evaluation tasks.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.task-normalized aggregate scoreDataset & 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 scrolls --batch_size autoopencompass --datasets scrolls --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 SCROLLS, 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