NeuroScore: 17,175 Texts Scored by Predicted Neural Engagement

The first public dataset of text content scored by predicted human brain activation using Meta's TRIBE v2 brain encoding model.

Each text has been processed through TRIBE v2, which predicts fMRI brain responses based on training data from 700+ human subjects. The output is predicted activation across 6 cortical regions of interest (ROIs), a weighted composite engagement score, and an emotion profile classification.

What's in the dataset

Field	Description
text	The raw text content
subject_line	Email subject line (for cold_email type, null otherwise)
content_type	One of: ad_copy, cold_email, sales_pitch, tagline, push_notification, linkedin_message
industry	Target industry (for generated texts) or "unknown" (for scraped)
emotional_strategy	Intended emotional framing (for generated texts) or "unknown"
source	"scraped" (from public datasets) or "generated" (via Anthropic API)
source_dataset	Original HuggingFace dataset name (for scraped texts)
reward_proxy	Predicted activation: ventromedial PFC / orbitofrontal (desire, value, "I want this")
attention_conflict	Predicted activation: anterior cingulate cortex (attention capture, surprise)
value_judgment	Predicted activation: frontopolar / medial PFC (trust, credibility, decisions)
social_semantic	Predicted activation: temporal pole / STS (social meaning, narrative)
language_processing	Predicted activation: Broca's + Wernicke's areas (linguistic engagement)
insula_activation	Predicted activation: anterior insula (gut feeling, empathy, visceral response)
composite_score	Weighted combination of ROI scores, 0 to 100 scale
emotion_profile	Classification based on dominant ROIs (see below)
brain_activation_mean	Mean activation across all cortical vertices
brain_activation_std	Standard deviation of activation
brain_activation_max	Peak activation value
temperature	Generation temperature (for AI-generated texts only)

Quick start

from datasets import load_dataset

ds = load_dataset("tushar710/neuroscore")
df = ds["train"].to_pandas()

# Highest scoring texts
print(df.nlargest(10, "composite_score")[["text", "composite_score", "emotion_profile"]])

# Average score by content type
print(df.groupby("content_type")["composite_score"].mean().sort_values(ascending=False))

# Average score by emotional strategy (generated texts only)
gen = df[df["emotional_strategy"] != "unknown"]
print(gen.groupby("emotional_strategy")["composite_score"].mean().sort_values(ascending=False))

How it was built

Data collection

17,175 texts from two sources:

Scraped (16,263 texts): Pulled from public HuggingFace datasets including PeterBrendan/Ads_Creative_Text_Programmatic, marketeam/Marketing-Emails, Yale-LILY/aeslc (Enron email corpus), PeterBrendan/AdImageNet, goendalf666/sales-conversations, RafaM97/marketing_social_media, google-research-datasets/go_emotions, dair-ai/emotion, christinacdl/clickbait_notclickbait_dataset, and others.

Generated (912 texts): Produced via the Anthropic API (Claude Sonnet) across a controlled matrix of 6 content types × 10 industries × 8 emotional strategies (fear_urgency, aspiration_reward, social_proof, empathy_pain, curiosity_intrigue, authority_trust, scarcity_exclusivity, humor_relatability).

Brain scoring

Each text was processed through TRIBE v2 (Meta FAIR, March 2026), a tri-modal foundation model trained on 700+ subjects' fMRI data.

Texts were scored using the language pathway (LLaMA 3.2 embeddings → transformer integration → cortical surface mapping). Predicted activations were extracted at ~20,484 vertices on the fsaverage5 cortical mesh, then aggregated into ROI scores using the Destrieux atlas parcellation.

Processing was performed on Google Colab (T4 GPU) with batch inference, averaging 3 to 15 seconds per text.

Composite score

Weighted combination of 6 working ROIs, normalized to 0 to 100:

composite = (0.25 × reward_proxy + 0.25 × attention_conflict + 0.20 × value_judgment 
           + 0.12 × social_semantic + 0.10 × language_processing + 0.08 × insula_activation) × 100

ROI scores are min-max normalized across the full dataset before weighting.

Emotion profiles

Classified by the top 2 dominant ROIs per text:

Profile	Dominant ROIs	Count
mixed	No clear dominant pair	8,362
fear-urgency	attention + insula	6,022
high-arousal-desire	attention + reward	1,521
curiosity-hook	attention + language	1,007
social-aspiration	reward + social	106
empathy-based	insula + social	88
reward-driven	reward + value	45
trust-authority	social + value	24

Key findings

Enron emails scored highest. Internal corporate emails about taxes, accounting, and pay stubs outscored all crafted content. Dense with numbers, carrying real stakes, demanding action. Composite scores up to 92.3.

Loss framing > gain framing. "You're losing X" triggered 9.2% more activation than "You could save X." Loss framing activated both attention and reward circuits; gain framing activated reward alone.

Specificity drives attention. Texts opening with a number activated 40% more attention-related cortex than texts opening with a question.

Longer text activates more brain. Sales pitches (62.0 mean) outscored taglines (43.4) and ad copy (40.4). More context = more simultaneous ROI activation.

Emotional strategy ranking. Scarcity (71.7) > curiosity (70.4) > fear (69.7) > humor (68.2) > social proof (67.9) > empathy (66.9) > authority (66.4) > aspiration (65.7).

Known limitations

Subcortical regions missing. TRIBE v2 maps cortical surface only. Amygdala (fear/emotional intensity) and hippocampus (memory encoding) are subcortical and returned all zeros. These were excluded from the dataset. The emotional picture is incomplete.

Predicted, not measured. TRIBE v2 predicts group-averaged fMRI responses. Individual brains vary. These are statistical tendencies across a population, not deterministic outcomes.

Activation ≠ effectiveness. High neural engagement could mean deep interest or strong aversion. The model does not distinguish valence in all regions.

Corpus bias. English-only. Commercially weighted (ads, emails, sales). Do not generalize to other languages, registers, or communication contexts.

Composite score is opinionated. The weights reflect a cognitive priority model (attention and reward weighted highest). Other weighting schemes are valid. Raw ROI scores are provided for custom scoring.

Licensing and attribution

This dataset is released under CC BY-NC-SA 4.0.

Brain scores were generated using TRIBE v2 by Meta FAIR, released under CC BY-NC. Generated texts were produced via the Anthropic API. Scraped texts originate from their respective public datasets on HuggingFace (see source_dataset field for attribution).

If you use this dataset, please cite:

@dataset{neuroscore2026,
  title={NeuroScore: Text Scored by Predicted Neural Engagement},
  author={Tushar Gautam},
  year={2026},
  url={https://huggingface.co/datasets/tushar710/neuroscore},
  note={Brain scores generated using Meta FAIR's TRIBE v2 model}
}

Contact

Built by Tushar Gautam, software engineer at Ampirial (Bangalore). Find me on LinkedIn to discuss or collaborate.