Datasets:

montagovian
/

basedBench

license: other
task_categories:
  - visual-question-answering
  - image-to-text
tags:
  - memes
  - vlm
  - benchmark
  - humor-understanding
dataset_info:
  - config_name: judgments
    features:
      - name: judgment_id
        dtype: int64
      - name: snapshot_id
        dtype: string
      - name: prediction_id
        dtype: int64
      - name: post_id
        dtype: string
      - name: model_id
        dtype: string
      - name: judge_model
        dtype: string
      - name: verdict
        dtype: string
      - name: reasoning
        dtype: string
      - name: judge_prompt_id
        dtype: string
      - name: judged_at
        dtype: string
      - name: is_latest
        dtype: bool
    splits:
      - name: train
        num_bytes: 7741795
        num_examples: 10660
    download_size: 2739183
    dataset_size: 7741795
  - config_name: leaderboard
    features:
      - name: snapshot_id
        dtype: string
      - name: model_id
        dtype: string
      - name: judge_model
        dtype: string
      - name: correct
        dtype: int64
      - name: incorrect
        dtype: int64
      - name: total
        dtype: int64
      - name: accuracy
        dtype: float64
      - name: judged_by_multiple
        dtype: int64
      - name: unanimous_agreements
        dtype: int64
      - name: agreement_rate
        dtype: float64
    splits:
      - name: train
        num_bytes: 543
        num_examples: 5
    download_size: 5148
    dataset_size: 543
  - config_name: memes
    features:
      - name: snapshot_id
        dtype: string
      - name: post_id
        dtype: string
      - name: title
        dtype: string
      - name: subreddit
        dtype: string
      - name: ground_truth
        dtype: string
      - name: image
        dtype: image
    splits:
      - name: train
        num_bytes: 170689057
        num_examples: 519
    download_size: 170630238
    dataset_size: 170689057
  - config_name: predictions
    features:
      - name: prediction_id
        dtype: int64
      - name: snapshot_id
        dtype: string
      - name: post_id
        dtype: string
      - name: model_id
        dtype: string
      - name: prediction
        dtype: string
      - name: dataset_version
        dtype: string
      - name: prediction_prompt_id
        dtype: string
      - name: latency_ms
        dtype: int64
      - name: token_count
        dtype: int64
      - name: created_at
        dtype: string
      - name: consensus_verdict
        dtype: string
      - name: judge_count
        dtype: int64
      - name: correct_votes
        dtype: int64
      - name: incorrect_votes
        dtype: int64
    splits:
      - name: train
        num_bytes: 4203146
        num_examples: 2785
    download_size: 2357404
    dataset_size: 4203146
configs:
  - config_name: judgments
    data_files:
      - split: train
        path: judgments/train-*
  - config_name: leaderboard
    data_files:
      - split: train
        path: leaderboard/train-*
  - config_name: memes
    data_files:
      - split: train
        path: memes/train-*
  - config_name: predictions
    data_files:
      - split: train
        path: predictions/train-*

BasedBench: basedBench-519-2026-07

BasedBench is a VLM meme-understanding benchmark. This snapshot contains 519 human-validated memes with ground-truth explanations derived from Reddit comment consensus.

Task Definition

The task is to determine whether a model gets the joke in a meme. A correct prediction identifies the relevant people, events, meme formats, media, phrases, visual details, or cultural references, then reconstructs the intended setup, implication, contrast, inversion, irony, wordplay, or other mechanism a viewer must notice to understand the meme.

This benchmark does not test whether a model can produce a psychological or aesthetic theory of why something is funny.

Leaderboard

Model	Correct	Total	Accuracy
claude-opus-4-8	312	519	60.1%
google/gemini-3.1-pro-preview	442	519	85.2%
gpt-5.5	419	519	80.7%
muse-spark-1.1	395	519	76.1%
x-ai/grok-4.3	254	519	48.9%

Dataset Contents

The public artifact includes:

Reddit post IDs, titles, and subreddit names.
Meme images used as benchmark stimuli.
Human-validated ground-truth explanations.
Every successful model prediction retained by BasedBench for this snapshot.
Every judge verdict and reasoning record for those predictions, including superseded rejudgments.
Derived consensus fields and leaderboard totals.

Raw Reddit comments, Reddit authors, review metadata, reviewer notes, consensus source comment IDs, local file paths, API request metadata, internal prompts, raw LLM responses, and LLM call logs are intentionally omitted.

Usage

from datasets import load_dataset

memes = load_dataset("montagovian/basedBench", "memes")
predictions = load_dataset("montagovian/basedBench", "predictions")
judgments = load_dataset("montagovian/basedBench", "judgments")
leaderboard = load_dataset("montagovian/basedBench", "leaderboard")

Dataset Structure

The dataset uses normalized long-form tables joined by snapshot_id, post_id, and prediction_id:

memes has one row per meme in the snapshot, with the image, title, subreddit, and human-validated ground truth.
predictions has one row per successful model prediction. It includes the target model, prediction text, dataset and prompt versions where available, latency, token count, timestamp, and derived consensus vote fields.
judgments has one row per judgment attempt. It includes the target and judge models, verdict, reasoning, judge prompt ID, timestamp, and an is_latest marker. Historical rejudgments remain present.
leaderboard has one derived row per target model, including score coverage and judge agreement statistics.

Failed API calls and operational error messages are not benchmark predictions and are omitted from the public tables. The current BasedBench database retains one successful prediction per (post_id, model_id) pair; historical successful prediction reruns that were never retained by the database cannot be exported.

Methodology

Ground-truth explanations are extracted from Reddit comments via LLM consensus detection. A candidate ground truth must reflect at least three substantive comments agreeing on the same specific explanation, and a human reviewer must validate the meme before it enters a release snapshot.

Predictor models receive only the meme image. They do not receive the Reddit title, comments, subreddit, ground truth, web search, or external tools.

Each model prediction is scored by an LLM judge ensemble using strict criteria: the judge asks whether the model recovered the same joke as the ground truth. For derived consensus fields and leaderboard accuracy, only the latest judgment from each (prediction_id, judge_model) pair is counted. A prediction receives a consensus verdict when at least two judges cast the same verdict and that verdict has a strict majority; otherwise it is omitted from the leaderboard denominator. All individual and historical judgments remain available in the judgments config.

License and Rights

This dataset has mixed rights status, so the machine-readable Hugging Face license is other.

Materials created and controlled by the BasedBench maintainers are released under the MIT License. This includes the benchmark code, dataset schema, export format, evaluation prompts where applicable, benchmark-specific metadata, leaderboard tables, judge verdicts and reasoning, and other maintainer-authored documentation and annotations to the extent the maintainers own or control those materials.

Meme images, Reddit post titles, subreddit names, post IDs, cultural references, logos, characters, screenshots, and other source artifacts may be owned by third parties. The BasedBench maintainers do not claim ownership of those underlying third-party materials. The MIT License for this repository does not apply to those third-party materials; all such rights remain with their respective owners.

The meme images and limited source metadata are included under a fair-use rationale for research, criticism, commentary, and benchmark evaluation. The use is transformative: the images are used as individual test stimuli for measuring whether vision-language models understand the intended joke, not as a substitute for the original posts or images. The dataset uses only the material needed to support that benchmark task, omits raw comment text and authors, and does not serve as a general-purpose meme archive or replacement market for the source works.

Users are responsible for determining whether their downstream use of any third-party material is permitted by law or by the relevant rights holder. If you believe specific material should not be included, please contact the dataset maintainers through the Hugging Face repository or the project repository.

Intended Use

This dataset is intended for research and evaluation of multimodal model understanding, especially whether models can connect visual details, text, cultural references, and joke structure.

It is not intended for training models to impersonate Reddit users, reconstruct deleted discussions, identify commenters, or build a general meme redistribution corpus.

Snapshot

Snapshot name: basedBench-519-2026-07
Snapshot ID: ad24bf870ce6285d
Created: 2026-07-15T01:49:13.506333+00:00
Memes: 519
Successful predictions: 2785
Judgment records: 10660