[ACL 2026] LaMPE: Length-aware Multi-grained Positional Encoding for Adaptive Long-context Scaling Without Training

LaMPE contains evaluation scripts for long-context position mapping experiments across RULER, InfiniteBench, LongBench, and LEval.

Layout

• ruler/: RULER data preparation and evaluation scripts.
• InfiniteBench/: InfiniteBench evaluation scripts and task data.
• longbench_eval/LongBench-main/: LongBench prediction and evaluation scripts.
• LEval/: LEval baselines, data, and evaluation scripts.
• run_all.sh: top-level launcher for the benchmark scripts.

Usage

Run one benchmark group:

bash run_all.sh ruler
bash run_all.sh infinitebench
bash run_all.sh longbench
bash run_all.sh leval

Run all benchmark groups in sequence:

bash run_all.sh all

Model paths, data paths, output directories, and Conda environments can be configured with environment variables such as LLAMA_MODEL_PATH, LONGBENCH_DATA_DIR, RULER_OUTPUT_DIR, CONDA_PROFILE, and LAMPE_CONDA_ENV.

Requirements

Install the tested core environment:

pip install -r requirements.txt

Key package versions are pinned from the development environment:

• torch==2.0.1
• transformers==4.43.3
• flash-attn==2.5.6
• accelerate==1.3.0
• datasets==3.6.0
• evaluate==0.4.3

Benchmark-specific requirements are also kept in InfiniteBench/requirements.txt and longbench_eval/LongBench-main/requirements.txt.

Data

Large benchmark inputs are not committed to this repository. Set INFINITEBENCH_DATA_DIR, LONGBENCH_DATA_DIR, and LONGBENCH_E_DATA_DIR to point to local copies of the corresponding benchmark data before running evaluations.