huaXiaKyrie/deltalora-memory-multisource-seed42

---
pretty_name: Delta-LoRA Memory Multisource Seed42
license: other
language:
- en
tags:
- long-context
- question-answering
- dialogue
- multi-hop
- temporal-reasoning
size_categories:
- 100K<n<1M
---

# Delta-LoRA Memory Multisource Seed42

This repository contains a mixed memory-training corpus built for Delta-LoRA long-context and online-memory experiments.

Files:

- `deltalora_memory_multisource_large_seed42.jsonl`: full mixed corpus
- `deltalora_memory_multisource_large_seed42.jsonl.summary.json`: summary for the full corpus
- `deltalora_memory_multisource_6k_seed42.jsonl`: stratified 6k subset for quick training
- `deltalora_memory_multisource_6k_seed42.jsonl.summary.json`: summary for the 6k subset

Source mixture in the full corpus:

- `convomem`: 16,000
- `hotpotqa`: 90,447
- `2wikimultihopqa`: 167,454
- `musique`: 39,876
- `quality`: 2,523
- `timeqa`: 28,989
- `squad_v2`: 130,319
- `quac`: 11,567
- `doc2dial`: 3,474

Total examples: `490,649`

The 6k subset is stratified as:

- `convomem`: 600
- `hotpotqa`: 600
- `2wikimultihopqa`: 1,200
- `musique`: 600
- `quality`: 600
- `timeqa`: 600
- `squad_v2`: 600
- `quac`: 600
- `doc2dial`: 600

Notes:

- `qasper` was intentionally excluded from this release.
- The dataset is normalized into a chat-style `messages` format for episode-style memory training.
- Licenses differ across original sources; please check the upstream datasets before reuse or redistribution.

Upstream sources used in this release:

- `Salesforce/ConvoMem`
- `hotpotqa/hotpot_qa`
- `ohjoonhee/2WikiMultihopQA`
- `voidful/MuSiQue`
- `emozilla/quality`
- `hugosousa/TimeQA`
- `rajpurkar/squad_v2`
- `allenai/quac`
- `IBM/doc2dial`