Datasets:

huaXiaKyrie
/

delta-mem-qasper-data

messages list	source_dataset string	source_split string	memory_skill_tags list	context_format string	answer_style string	num_context_chunks int64	num_annotations int64	paper_id string	question_id string	prompt_profile string	episode_safe bool	context_token_count int64	full_chat_token_count int64	episode_visible_message_count int64
[ { "role": "system", "content": "You answer questions using earlier notes, passages, or dialogue history. Use exact details from them whenever possible. If they do not contain the answer, reply exactly: No information available." }, { "role": "user", "content": "History turn 1 (User):\nPaper chun...	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	34	1	1909.06937	b4f5bf3b7b37e2f22d13b724ca8fe7d0888e04a2	episode_flattened_v1	true	6,349	6,407	2
[ { "role": "system", "content": "You answer questions using earlier notes, passages, or dialogue history. Use exact details from them whenever possible. If they do not contain the answer, reply exactly: No information available." }, { "role": "user", "content": "History turn 1 (User):\nPaper chun...	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	23	1	1911.03243	8847f2c676193189a0f9c0fe3b86b05b5657b76a	episode_flattened_v1	true	3,617	3,671	2
[ { "role": "system", "content": "You answer questions using earlier notes, passages, or dialogue history. Use exact details from them whenever possible. If they do not contain the answer, reply exactly: No information available." }, { "role": "user", "content": "History turn 1 (User):\nPaper chun...	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	29	1	2003.09586	e75f5bd7cc7107f10412d61e3202a74b082b0934	episode_flattened_v1	true	5,902	5,988	2
[ { "role": "system", "content": "You answer questions using earlier notes, passages, or dialogue history. Use exact details from them whenever possible. If they do not contain the answer, reply exactly: No information available." }, { "role": "user", "content": "History turn 1 (User):\nPaper chun...	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	18	1	1912.01673	016b59daa84269a93ce821070f4f5c1a71752a8a	episode_flattened_v1	true	3,774	3,835	2
[ { "role": "system", "content": "You answer questions using earlier notes, passages, or dialogue history. Use exact details from them whenever possible. If they do not contain the answer, reply exactly: No information available." }, { "role": "user", "content": "History turn 1 (User):\nPaper chun...	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	27	1	1709.05404	c5ac07528cf99d353413c9d9ea61a1a699dd783e	episode_flattened_v1	true	6,581	6,644	2
[ { "role": "system", "content": "You answer questions using earlier notes, passages, or dialogue history. Use exact details from them whenever possible. If they do not contain the answer, reply exactly: No information available." }, { "role": "user", "content": "History turn 1 (User):\nPaper chun...	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	31	1	2002.01030	144714fe0d5a2bb7e21a7bf50df39d790ff12916	episode_flattened_v1	true	5,805	5,876	2
[ { "role": "system", "content": "You answer questions using earlier notes, passages, or dialogue history. Use exact details from them whenever possible. If they do not contain the answer, reply exactly: No information available." }, { "role": "user", "content": "History turn 1 (User):\nPaper chun...	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	22	1	1907.08501	9653c89a93ac5c717a0a26cf80e9aa98a5ccf910	episode_flattened_v1	true	4,288	4,370	2
[ { "role": "system", "content": "You answer questions using earlier notes, passages, or dialogue history. Use exact details from them whenever possible. If they do not contain the answer, reply exactly: No information available." }, { "role": "user", "content": "History turn 1 (User):\nPaper chun...	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	16	1	1907.11907	8c981f8b992cb583e598f71741c322f522c6d2ad	episode_flattened_v1	true	4,037	4,103	2
[{"role":"system","content":"You answer questions using earlier notes, passages, or dialogue history(...TRUNCATED)	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	17	1	1811.00266	013a8525dbf7a9e1e69acc1cff18bb7b8261cbad	episode_flattened_v1	true	3,343	3,394	2
[{"role":"system","content":"You answer questions using earlier notes, passages, or dialogue history(...TRUNCATED)	qasper	train	[ "long_document_reasoning", "academic_document_qa" ]	document	span_yes_no_or_short_phrase	36	1	1909.05855	3ee721c3531bf1b9a1356a40205d088c9a7a44fc	episode_flattened_v1	true	6,642	6,723	2

End of preview. Expand in Data Studio

Introduction

This repository contains the δ-mem training data, as presented in the paper δ-mem: Efficient Online Memory for Large Language Models.

δ-mem is a lightweight online memory mechanism that augments a frozen backbone with a compact associative memory state. It projects context into a low-dimensional space and updates a state matrix via delta-rule learning, allowing for efficient long-term memory utilization without full fine-tuning or context extension.

Paper: https://huggingface.co/papers/2605.12357 Repository: https://github.com/declare-lab/delta-Mem

Citation

@misc{lei2026deltamem,
  title         = {$\delta$-mem: Efficient Online Memory for Large Language Models},
  author        = {Jingdi Lei and Di Zhang and Junxian Li and Weida Wang and Kaixuan Fan and Xiang Liu and Qihan Liu and Xiaoteng Ma and Baian Chen and Soujanya Poria},
  year          = {2026},
  eprint        = {2605.12357},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2605.12357}
}

Downloads last month: 61

Paper for huaXiaKyrie/delta-mem-qasper-data

δ-mem: Efficient Online Memory for Large Language Models

Paper • 2605.12357 • Published 23 days ago • 125