ZeroSCROLLS
Zero-shot long-text suite derived from SCROLLS with test-only tasks and new aggregation-style long-context challenges.
ZeroSCROLLS fixed SCROLLS' fine-tuning mismatch and has a live leaderboard, but its 2023-era tasks are now shorter and narrower than current 100k+ and multi-needle long-context stress suites.
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 |
|---|---|---|---|---|
| T0pp (paper Table 3 Avg) | 14.3% | 2023-05-23 | independent | Source ↗ |
| Flan-T5 (paper Table 3 Avg) | 29.9% | 2023-05-23 | independent | Source ↗ |
| Flan-UL2 (paper Table 3 Avg) | 30.6% | 2023-05-23 | independent | Source ↗ |
| DaVinci003 (paper Table 3 Avg) | 33.7% | 2023-05-23 | independent | Source ↗ |
| ChatGPT (paper Table 3 Avg) | 34% | 2023-05-23 | independent | Source ↗ |
| Claude (paper Table 3 Avg) | 39.1% | 2023-05-23 | independent | Source ↗ |
| GPT-4 (paper Table 3 Avg) | 41.7% | 2023-05-23 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.No single suite-wide human baseline is recorded here; component tasks have their own annotation histories.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.task-normalized zero-shot aggregateDataset & 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 zeroscrolls --batch_size autoopencompass --datasets zeroscrolls --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 ZeroSCROLLS, 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