apoorvumang's picture
Add knowledge-cutoff benchmark: 330 events + 7920 results (12 models)
70ac833 verified
|
Raw
History Blame Contribute Delete
3.81 kB
metadata
license: cc-by-4.0
language:
  - en
task_categories:
  - question-answering
tags:
  - knowledge-cutoff
  - llm-evaluation
  - temporal-reasoning
  - benchmark
pretty_name: Knowledge Cutoff Benchmark
size_categories:
  - 1K<n<10K
configs:
  - config_name: events
    default: true
    data_files: events.jsonl
  - config_name: results
    data_files: results.jsonl

Knowledge Cutoff Benchmark

A benchmark for estimating a language model's effective knowledge cutoff — what it actually knows about the world — which is usually earlier than the cutoff date the model advertises.

Each model is probed on curated, surprising / unforecastable real-world events (deaths, changes of office) spread month-by-month across Jan 2024 – Jun 2026. The month where per-month accuracy collapses is the model's effective knowledge horizon.

Configs

events (default) — the benchmark items (one row per event)

field meaning
id unique event id
date, month when it happened (YYYY-MM-DD, YYYY-MM)
category death, office_change, control_alive, fake_event
predictability low = unforecastable (highest signal) → high
region US / International
subject, fact who/what, and a one-sentence ground truth
question_direct, expected_direct open probe + canonical answer
mcq_question, mcq_choices, mcq_answer 4-way forced-choice probe
source provenance URL

control_alive (a famous person still living) and fake_event (an event that never happened) are diagnostics: a trustworthy run answers them correctly, which rules out a model that just always guesses "dead" or confabulates.

results — model answers (one row per model × probe × event)

field meaning
model, probe model key and direct (open) or mcq (forced choice)
event_id, month, category, predictability joins back to events
label correct / incorrect (confidently wrong) / abstain
response the model's raw answer

Method (why the numbers are trustworthy)

  1. Surprising events only — a model can't score by extrapolating pre-cutoff trends the way it could for a scheduled election.
  2. Three-way gradingcorrect / incorrect / abstain. Hedging ("I'm not aware…") is abstention, not error; conflating them would corrupt the estimate. (Direct answers are graded by an LLM judge given the ground truth, so the judge's own cutoff is irrelevant; MCQ is graded by letter.)
  3. Controls — living-person and fabricated-event rows that every run must pass.
  4. No date leakage — prompts never reveal the current date.

direct under-counts (the model knows but doesn't volunteer); mcq over-counts (guessing) — together they bracket the truth.

Key finding

Across the evaluated frontier models, the effective knowledge cutoff is consistently ~1–5 months earlier than the advertised cutoff. Models that publish no cutoff (e.g. some open-weight releases) can only be characterized by this benchmark. See the live explorer for the per-model leaderboard, decay curves, a heatmap with each model's claimed cutoff highlighted, and every individual answer.

Usage

from datasets import load_dataset

events  = load_dataset("apoorvumang/knowledge-cutoff-benchmark", "events",  split="train")
results = load_dataset("apoorvumang/knowledge-cutoff-benchmark", "results", split="train")

License

cc-by-4.0. Events are factual and drawn from public reporting (see each row's source); please cite this dataset if you use it.