BEAM Dataset

July 24, 2026 · View on GitHub

User Memory Evaluation guide

BEAM (Beyond a Million Tokens) is a long-term memory benchmark designed for extreme-length scenarios, covering four scales from 128K to 10M tokens. Published by Université de Montréal et al. at ICLR 2026.

Files

FileSizeConversationsQuestionsDescription
beam_100k.json~14.1 MB20400128K tokens scale
beam_500k.json~85.9 MB35700500K tokens scale
beam_1m.json~172.3 MB357001M tokens scale
beam_10m_10m.json~979.5 MB1020010M tokens scale

Only README.md and prepare_beam.py are intended to be version-controlled by OmniMemEval. Downloaded JSONL files are generated artifacts and should remain local.

Dataset Statistics

MetricValue
Total conversations100
Total questions2,000
Scale levels4 (128K / 500K / 1M / 10M tokens)
Memory capability dimensions10
Questions per conversation20 (2 per dimension)

10 Memory Capability Dimensions

DimensionQuestions/ConversationDescription
abstention2Identify questions unanswerable from conversation
contradiction_resolution2Detect and resolve contradictory information
event_ordering2Correctly order events chronologically
information_extraction2Extract specific facts from conversation
instruction_following2Follow instructions given in conversation
knowledge_update2Track information updates over time
multi_session_reasoning2Integrate information across sessions
preference_following2Identify and follow user preferences
summarization2Accurately summarize conversation content
temporal_reasoning2Reason about temporal relationships

Data Structure

JSON Lines format. Each conversation object contains:

FieldTypeDescription
conversation_idstringUnique conversation identifier
conversation_seeddictSeed info with category (Coding/General/Math), subtopics
narrativesstringConversation narrative outline
user_profiledictUser profile information
conversation_planstringConversation plan
user_questionslistUser questions with messages and time_anchor
chatlistFull dialogue with content, role, time_anchor, question_type per turn
probing_questionsstring20 test questions (JSON string) organized by 10 dimensions

Data Preparation

The JSON Lines files are converted from the original Hugging Face Parquet datasets. The default download matches the OmniMemEval runner default (--scale 100k) so that a new developer does not accidentally fetch the full 1.3 GB converted dataset.

pip install datasets
python data/beam/prepare_beam.py                   # download 100k (~14 MB)
python data/beam/prepare_beam.py --scale 100k      # download a single scale
python data/beam/prepare_beam.py --scale 100k 500k # download selected scales
python data/beam/prepare_beam.py --scale all       # download all scales
python data/beam/prepare_beam.py --force           # overwrite existing files

Note: The 10M scale file is ~979 MB after JSONL conversion. Use --scale to selectively download smaller scales first.

License And Redistribution

The BEAM datasets are released under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0). Redistribution or adapted copies must preserve attribution and comply with the ShareAlike terms. The OmniMemEval code license does not override the upstream BEAM dataset license.

Evaluation

  • Metric: Nugget Score (LLM-as-a-Judge), scored per atomic nugget (1.0 / 0.5 / 0.0)
  • Pipeline: Ingest conversations → Search memories → Answer probing questions → LLM judge scores against rubric
  • Scale-stratified reporting: Separate scores for 128K / 500K / 1M / 10M

References