GACC-Bench

Benchmarking Goal-Adaptive Context Compression Ability across Large Language Models

Paper | Code | Dataset

Overview

GACC-Bench evaluates whether a language model can retain information relevant to a specified goal under a context-length budget. The compressor sees a source context, a goal, and a token budget. A reader then answers the associated questions using the compressed context; those questions are withheld from the compressor.

This Hugging Face release preserves the goal-level records and nested questions from the GitHub release. Parquet files support datasets.load_dataset and the dataset viewer. Original JSONL files remain available under original/ for the evaluator.

Dataset statistics

Configuration	Split	Contexts	Goals / rows	Questions
default	test	281	500	2,571
testmini	test	62	124	651

testmini is an exact subset of the complete benchmark, with two goals per selected context. It is a convenience configuration for smaller evaluations, not a disjoint validation set. Both configurations are evaluation data.

Source collection	Contexts	Goals	Questions
HELMET	27	40	208
L-Eval (LEval)	67	102	526
LongBench-v2	23	49	246
LongCodeBench	35	78	391
Loong	44	102	573
MeetingBank	67	101	489
SWE-agent	18	28	138

Load with datasets

from datasets import load_dataset

dataset = load_dataset("Regimes/GACC-Bench", split="test")
mini = load_dataset("Regimes/GACC-Bench", name="testmini", split="test")

example = dataset[0]
print(example["dataset"], example["goal"])
print(example["questions"][0]["question"])

Each row contains one goal. The same source context can appear in multiple rows with different goals.

Field	Type	Description
dataset	string	Source collection identifier
context_id	string	Source context identifier
context	string	Full source context
goal_id	string	Unique goal identifier
goal	string	High-level information-retention goal
questions	list of objects	Questions associated with this goal
questions[].question_id	string	Question identifier
questions[].question	string	Evaluation question
questions[].gold_answer	string	Reference answer

Use with the evaluation code

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Regimes/GACC-Bench",
    repo_type="dataset",
    filename="original/GACC-Bench.jsonl",
    local_dir="hf_data",
)
print(path)

Pass the downloaded file to run.sh --input in the evaluation repository. The default compression budgets are 5% and 10% of source-context tokens. Code, prompts, metrics, and command-line options are documented there.

Provenance

SOURCE.json records the source commit, byte-level SHA-256 checksums of the original JSONL files, and checksums of the Parquet exports. The Parquet data retains the original row order, field values, and question order. Contexts originate from the seven source collections listed above; retain their original attribution when reusing those materials.

Citation

@misc{liu2026gaccbench,
  title={GACC-Bench: Benchmarking Goal-Adaptive Context Compression Ability across Large Language Models},
  author={Weiting Liu and Zhiyu Pan and Shuhong Wang and Zhenquan Yao and Jungang Li and Li Zhao and Jian Jiang and Jianfeng Feng and Ai Han and Wenlian Lu},
  year={2026},
  eprint={2610.21893},
  archivePrefix={arXiv},
  primaryClass={cs.IR},
  url={https://arxiv.org/abs/2610.21893}
}