zhangyir/Copy_Benchmark

Copy Benchmark

This benchmark evaluates the copying ability of language models and consists of three copy evaluation subsets. These subsets correspond to the three test sets used in the subfigures of Figure 1 in our paper (link coming soon). The benchmark contains sequences with a lot of repetitive patterns, which are difficult for large language models to copy. As shown in our paper, even frontier large language models struggle on these benchmarks. A detailed description of the benchmark construction and data generation process can be found in Section A of the paper. Notice that the dataset isn't used to do finetuning in our LLM experiments.

subset	examples	model input	scoring
binary-copy-recursive-flip	600	binary strings using 0/1	exact string match
binary-copy-imbalanced	300	spaced symbols using a / b	extract a/b/A/B from model output, map A -> a, B -> b, compare with the target
python-list-conversion	200	comma-separated numeric sequence	extract numbers and compare the full numeric sequence

All subsets have the same top-level columns:

id, input, metadata

The prompt and gold answer are stored in:

record["input"]["messages"]

The first assistant message is the gold answer.

Generate the benchmark files

python generate_benchmark.py --output-dir .

This writes:

data/binary-copy-recursive-flip.jsonl
data/binary-copy-imbalanced.jsonl
data/python-list-conversion.jsonl
benchmark_manifest.json

Load from Hugging Face

from datasets import load_dataset

ds = load_dataset("zhangyir/Copy_Benchmark", "binary-copy-imbalanced", split="train")
print(ds[0]["input"]["messages"])

Evaluate a Hugging Face model

Install dependencies:

pip install -U datasets transformers accelerate torch tqdm

Run one subset:

python eval_benchmark.py \
  --model Qwen/Qwen2.5-7B-Instruct \
  --dataset zhangyir/Copy_Benchmark \
  --subset binary-copy-imbalanced

For local JSONL files:

python eval_benchmark.py \
  --model Qwen/Qwen2.5-7B-Instruct \
  --data-file data/binary-copy-imbalanced.jsonl \
  --subset binary-copy-imbalanced

For testing local models:

python eval_benchmark.py \
  --model /path/to/local/model \
  --data-file data/binary-copy-imbalanced.jsonl \
  --subset binary-copy-imbalanced

For a quick sanity check, use --limit to evaluate only the first few rows of the selected subset. For example, --limit 5 evaluates the first 5 rows only.

python eval_benchmark.py \
  --model Qwen/Qwen2.5-7B-Instruct \
  --dataset zhangyir/Copy_Benchmark \
  --subset python-list-conversion \
  --limit 5

Results are written to:

hf_eval_outputs/<model>/<subset>/predictions.jsonl
hf_eval_outputs/<model>/<subset>/summary.json

How to Cite This Paper?

@inproceedings{wen2026frontier,
    title={Frontier Language Models Struggle to Copy: Text Can Be Better Viewed in 2D},
    author={Haodong Wen and Yiran Zhang and Yingfa Chen and Kaifeng Lyu},
    booktitle={ICML 2026 Workshop on Structured Probabilistic Inference {\&} Generative Modeling},
    year={2026},
    url={https://openreview.net/forum?id=r3UVsj13Mr}
}