Datasetpublic

cozzyde/long-horizon-memorization

ComposeCL Datasets This repository contains the three 100-task memorization datasets from the paper Continual Learning Mechanisms Compose for Long-Horizon Memorization (arXiv: 2609.06986). Project page: https://compose-cl.github.io/ Code: https://github.com/cozheyuanzhangde/compose-cl Datasets Dataset Path Tasks Items/task Construction Symbol-QA data/synthetic_qa/symbol_qa 100 100 random six-character keys mapped to random four-character values… See the full description on the dataset page: https://huggingface.co/datasets/cozzyde/long-horizon-memorization.

sourceHugging Facemitupdated 24d agoView on Hugging Face

3likes198downloads

Dataset Card

ComposeCL Datasets

This repository contains the three 100-task memorization datasets from the paper Continual Learning Mechanisms Compose for Long-Horizon Memorization (arXiv: 2609.06986).

—Project page: https://compose-cl.github.io/
—Code: https://github.com/cozheyuanzhangde/compose-cl

Datasets

Dataset	Path	Tasks	Items/task	Construction
Symbol-QA	`data/synthetic_qa/symbol_qa`	100	100	random six-character keys mapped to random four-character values
LLM-QA	`data/synthetic_qa/llm_qa`	100	100	LLM-generated fictional facts, with task order shuffled using seed 0
Real-QA	`data/real_qa/qwen3_4b_base`	100	50	ten public QA sources, filtered against Qwen3-4B-Base and globally remixed

Training and evaluation use the same associations because the target quantity is memorization retention rather than held-out generalization. Answers are rendered as \boxed{answer} at load time. The stored JSON remains plain text.

The included Real-QA questions were filtered against Qwen/Qwen3-4B-Base and should be paired with that backbone when reproducing the paper.