Skip to content

Dataset · memorilla

Data with pre-computed embeddings for LLM memory modules

memorilla/Memorilla-Data

Millions of question-answer pairs with document collections and pre-computed Qwen3-Embedding-4B vectors.

What's inside

Contains English text data ranging from 1 to 10 million entries. Each row links a question and a reference answer with a set of source documents upon which the answer is based. The dataset was created to train and evaluate Memorilla, a memory module that compresses documents into a set of tokens for a frozen language model. Pre-calculated Qwen3-Embedding-4B vectors for all texts and questions are included.

What you can do with it

  • Training context compression and external memory modules for LLMs
  • Testing RAG systems using a specified document collection
  • Training models to answer questions from multiple sources without recomputing embeddings
Tasks
Question answering, Summarization, Text generation
Languages
English
Size
1M–10M rows
Downloads this month
1.8K
Likes
4

How to load it

from datasets import load_dataset

ds = load_dataset("memorilla/Memorilla-Data")