Memorilla data

Training and evaluation data for Memorilla, a memory module that compresses a collection of documents into a small set of memory tokens for a frozen language model. Every row pairs a question and a reference answer with the document collection the answer should be grounded in. Precomputed Qwen3-Embedding-4B vectors for every document and every question are included, so training and evaluation start without an embedding pass.

Usage

from datasets import load_dataset

pv4 = load_dataset("MemoryAsModality/Memorilla", "pv4", split="test")
row = pv4[0]
print(row["question"], row["answer"], len(row["documents"]))

The Memorilla code downloads the embeddings of a config on first use. To fetch them directly:

import torch
from huggingface_hub import snapshot_download
from safetensors import safe_open

config, split = "pv4", "test"
root = snapshot_download(
    "MemoryAsModality/Memorilla",
    repo_type="dataset",
    allow_patterns=[f"embeddings/qwen3-embedding-4b/{config}/{split}/*"],
)
folder = f"{root}/embeddings/qwen3-embedding-4b/{config}/{split}"

with safe_open(f"{folder}/index.safetensors", framework="pt") as f:
    index = {key: f.get_tensor(key) for key in f.keys()}

def documents(collection_id: int):
    position = int(torch.searchsorted(index["collection_ids"], torch.tensor(collection_id)))
    shard, start, length = (int(index[k][position]) for k in ("shard", "start", "length"))
    with safe_open(f"{folder}/documents-{shard:05d}.safetensors", framework="pt") as f:
        return f.get_slice("embeddings")[start : start + length]  # [num_documents, 2560]

with safe_open(f"{folder}/questions.safetensors", framework="pt") as f:
    questions = f.get_tensor("embeddings")  # [num_rows, 2560], row-aligned with the split

print(documents(row["collection_id"]).shape, questions[0].shape)

Configs

Benchmarks

config	dataset	rows	collection	embeddings
pv4	PersonalizationV4	train 15,058 · validation 1,755 · test 4,237	user	364.5 MiB
pmv2	PersonaMem-v2	train 18,521 · validation 2,059 · test 5,000	persona	1.1 GiB
factkg	FactKG	train 86,367 · validation 13,266 · test 9,041	claim	69.4 GiB
triviaqa	TriviaQA	train 138,384 · validation 17,444 · test 500	question	124.6 GiB
narrativeqa	NarrativeQA	train 32,747 · validation 3,461 · test 10,557	story	11.5 GiB
pubmedqa	PubMedQA	train 900 · test 100	question	23.9 MiB
lamp4	LaMP-4	train 12,527 · test 1,925	question	19.1 GiB
lamp7	LaMP-7	train 10,437 · test 1,500	question	1.1 GiB

Instruction mixture

config	dataset	rows	collection	embeddings
squad_v2	SQuAD 2.0	train 86,821 · validation 5,928	question	1.2 GiB
drop	DROP	train 77,400 · validation 9,535	question	1.5 GiB
coqa	CoQA	train 108,647 · validation 7,983	question	2.5 GiB
quail	QuAIL	train 10,246 · validation 2,164	question	313.9 MiB
pwc	PwC	train 241,564 · validation 18,141	question	6.9 GiB
cnn_dailymail	CNN/DailyMail	train 287,113 · validation 13,368	article	14.9 GiB
samsum	SAMSum	train 14,731 · validation 818	conversation	190.5 MiB
dialogsum	DialogSum	train 12,460 · validation 500	dialogue	174.9 MiB
msmarco	MS MARCO	train 504,111 · validation 55,597	query	30.7 GiB

Pretraining corpus

config	dataset	rows	collection	embeddings
enwiki	English Wikipedia	train 1,979,809 · validation 10,017	passage	26.0 GiB

• pv4 (PersonalizationV4). A fictional user's 200 conversations with an assistant and a scenario question about what the user would most likely do or prefer. train and validation hold the training and held-out questions of 119 users; test holds every question of 30 further users, none of whom appear in train or validation. hard marks 629 test questions that Qwen3-4B-Instruct-2507 gets wrong either with no documents or with only the single most relevant conversation.
• pmv2 (PersonaMem-v2). A simulated user's conversation history and a request whose best response depends on the user's preferences. Splits follow the per-persona files of the source: test holds 5,000 questions from 200 personas.
• factkg (FactKG). A claim to label True or False together with DBpedia triples about the claim's entities, written as subject → relation → object. Original train, dev (validation) and test claims.
• triviaqa (TriviaQA). A trivia question with the evidence documents of TriviaQA's reading-comprehension release (Wikipedia pages, or web search results when a question has none), split into passages. test is a fixed sample of 500 questions from the original validation set; validation holds the remaining 17,444. The original test questions are not included because their answers are not public.
• narrativeqa (NarrativeQA). A question about a book or movie script; the documents are the full story text split into chunks. Original train, validation and test splits.
• pubmedqa (PubMedQA). A biomedical research question over the sections of a PubMed abstract; the answer starts with yes, no or maybe followed by the long answer. The 1,000 expert-labeled questions (pqa_labeled), split 900 / 100 into train and test.
• lamp4 (LaMP-4). Personalized news headline generation: an article to title, with the author's previous articles as documents. train and test are the user-based LaMP-4 train and development sets.
• lamp7 (LaMP-7). Personalized tweet paraphrasing: a tweet to paraphrase, with the user's previous tweets as documents. train and test are the user-based LaMP-7 train and development sets.
• squad_v2 (SQuAD 2.0). A question over a Wikipedia paragraph; only answerable questions are included. Original train and validation splits.
• drop (DROP). A question requiring discrete reasoning (counting, arithmetic, sorting) over a paragraph. Original train and validation splits.
• coqa (CoQA). A conversational question over a passage; the question includes the preceding turns. Original train and validation splits.
• quail (QuAIL). A multiple-choice reading comprehension question (choices included in the question) over a passage; the answer is the text of the correct choice. Original train and validation splits.
• pwc (PwC). An instruction about a text (questions, summaries, extraction) with the text as documents. train is the original train split; validation is the original test split.
• cnn_dailymail (CNN/DailyMail). A news article to summarize; the answer is the article's highlights. Original train and validation splits (version 3.0.0).
• samsum (SAMSum). A messenger-style conversation to summarize. Original train and validation splits.
• dialogsum (DialogSum). A spoken-style dialogue to summarize. Original train and validation splits.
• msmarco (MS MARCO). A web search query with its retrieved passages and a human-written answer (the well-formed answer when there is one); queries without an answer are not included. Original train and validation splits (v2.1).
• enwiki (English Wikipedia). A Wikipedia passage, headed by its article title and section and split into chunks of at most about 128 tokens, with a generic instruction to restate it; the answer is the passage text. Train and validation passages.

Fields

field	type	description
collection_id	int64	Identifies the row's document collection. Rows with the same collection_id in a split share the same documents.
question	string	The question or instruction.
answer	string	The reference answer.
choices	list[string]	pv4, pmv2 and factkg only. Candidate answers for embedding-based scoring; answer is one of them. pv4 and pmv2 list five candidates in A-E order, with an empty string for a missing one; factkg lists ["True", "False"].
hard	bool	pv4 only. Marks the hard test subset (always False in train and validation).
documents	list[string]	The document collection, in its original order.

Embeddings

embeddings/qwen3-embedding-4b/<config>/<split>/
    index.safetensors            collection_ids int64 [C] (ascending), shard int32 [C], start int64 [C], length int64 [C]
    documents-XXXXX.safetensors  embeddings float16 [n, 2560]; whole collections, at most 4 GiB per shard
    questions.safetensors        embeddings float16 [R, 2560], row-aligned with data/<config>/<split>

• Collection index["collection_ids"][i] occupies rows start[i] to start[i] + length[i] of documents-{shard[i]:05d}.safetensors, one row per entry of documents, in order.
• Each split stores the collections it uses, so a collection shared by two splits appears in both.
• Vectors were computed with Qwen/Qwen3-Embedding-4B in vLLM (bfloat16, default pooling, unit-normalized) without an instruction prefix and stored as float16. Questions are embedded from question.strip().

Licensing

Each config is distributed under the license of its source dataset, listed below. Some sources are restricted to non-commercial use.

config	source	license	reference
pv4	MemoryAsModality/PersonalizationV4; persona seeds from Synthetic-Persona-Chat	CC BY 4.0	Memorilla; seed personas: Jandaghi et al., 2023. Faithful Persona-based Conversational Dataset Generation with Large Language Models.
pmv2	bowen-upenn/PersonaMem-v2	CC BY 4.0	Jiang et al., 2025. PersonaMem-v2: Towards Personalized Intelligence via Learning Implicit User Personas and Agentic Memory.
factkg	FactKG; triples from DBpedia	FactKG: no license stated; DBpedia: CC BY-SA 3.0	Kim et al., 2023. FactKG: Fact Verification via Reasoning on Knowledge Graphs. ACL.
triviaqa	mandarjoshi/trivia_qa (rc)	Apache 2.0	Joshi et al., 2017. TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension. ACL.
narrativeqa	deepmind/narrativeqa	Apache 2.0	Kočiský et al., 2018. The NarrativeQA Reading Comprehension Challenge. TACL.
pubmedqa	qiaojin/PubMedQA (pqa_labeled)	MIT	Jin et al., 2019. PubMedQA: A Dataset for Biomedical Research Question Answering. EMNLP.
lamp4	LaMP	See the LaMP benchmark	Salemi et al., 2024. LaMP: When Large Language Models Meet Personalization. ACL.
lamp7	LaMP	See the LaMP benchmark	Salemi et al., 2024. LaMP: When Large Language Models Meet Personalization. ACL.
squad_v2	rajpurkar/squad_v2	CC BY-SA 4.0	Rajpurkar et al., 2018. Know What You Don't Know: Unanswerable Questions for SQuAD. ACL.
drop	ucinlp/drop	CC BY-SA 4.0	Dua et al., 2019. DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs. NAACL.
coqa	stanfordnlp/coqa	Per-domain licenses of CoQA (CC BY-SA 4.0, MSR-LA, RACE, Apache 2.0)	Reddy et al., 2019. CoQA: A Conversational Question Answering Challenge. TACL.
quail	textmachinelab/quail	CC BY-NC-SA 4.0	Rogers et al., 2020. Getting Closer to AI Complete Question Answering: A Set of Prerequisite Real Tasks. AAAI.
pwc	sggetao/PwC	Apache 2.0	Ge et al., 2024. In-context Autoencoder for Context Compression in a Large Language Model. ICLR.
cnn_dailymail	abisee/cnn_dailymail	Apache 2.0	See et al., 2017. Get To The Point: Summarization with Pointer-Generator Networks. ACL.
samsum	knkarthick/samsum	CC BY-NC-ND 4.0	Gliwa et al., 2019. SAMSum Corpus: A Human-annotated Dialogue Dataset for Abstractive Summarization.
dialogsum	knkarthick/dialogsum	CC BY-NC-SA 4.0	Chen et al., 2021. DialogSum: A Real-Life Scenario Dialogue Summarization Dataset. Findings of ACL.
msmarco	microsoft/ms_marco (v2.1)	MS MARCO terms (non-commercial research use)	Nguyen et al., 2016. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset.
enwiki	English Wikipedia (December 2021 dump)	CC BY-SA 4.0	Wikipedia contributors.

Citation

If you use this data, please cite Memorilla (see the code repository) and the original dataset of each config you use (table above).