LongBench
Bilingual multitask suite for long-context understanding across QA, summarization, few-shot learning, synthetic tasks, and code completion.
LongBench remains a common baseline suite, but its original 5k-15k-heavy, public static task mix has been overtaken for frontier use by LongBench v2/Pro, RULER, MRCR, and 100k+ 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 |
|---|---|---|---|---|
| GPT-3.5-Turbo-16k (paper Table 3 OverAll) | 44.7% | 2023-08-28 | independent | Source ↗ |
| Llama2-7B-chat-4k (paper Table 3 OverAll) | 26.8% | 2023-08-28 | independent | Source ↗ |
| LongChat-v1.5-7B-32k (paper Table 3 OverAll) | 31.6% | 2023-08-28 | independent | Source ↗ |
| XGen-7B-8k (paper Table 3 OverAll) | 25% | 2023-08-28 | independent | Source ↗ |
| InternLM-7B-8k (paper Table 3 OverAll) | 22.6% | 2023-08-28 | independent | Source ↗ |
| ChatGLM2-6B (paper Table 3 OverAll) | 25.7% | 2023-08-28 | independent | Source ↗ |
| ChatGLM2-6B-32k (paper Table 3 OverAll) | 41.4% | 2023-08-28 | independent | Source ↗ |
| Vicuna-v1.5-7B-16k (paper Table 3 OverAll) | 30.5% | 2023-08-28 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Measured human performance on 16 long-context retrieval, QA, and summarization tasks.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.average score across task-specific metrics (%)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 longbench --batch_size autoopencompass --datasets longbench --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 LongBench, 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