JosefGoldstein/aimlessinnovations_customer_sentiment_v2

Aimless Innovations — Synthetic Sentiment Dataset v2 (AI vs. Karen Demo)

Synthetic e-commerce reviews and support messages for the "Aimless Innovations" website, used in the "AI vs. Karen" real-time sentiment analysis demo.

• Task: 6-class sentiment classification
• Classes: karen, very_negative, negative, neutral, positive, very_positive
• Generation: GPT-4.1 (synthetic content only)
• Format: JSONL with fields { id, text, label_text, label }
  • label_text ∈ {karen, very_negative, negative, neutral, positive, very_positive}
  • label ∈ {0, 1, 2, 3, 4, 5} aligned with the above order

Classes

Label	Class	Description
0	karen	Hostile, entitled, threatening, demands manager
1	very_negative	Strong dissatisfaction, frustration, anger
2	negative	Complaints, disappointment (not aggressive)
3	neutral	Questions, inquiries, factual statements
4	positive	Generally satisfied, mild praise
5	very_positive	Enthusiastic, glowing feedback

Splits and Files

Split	File	Examples
train	train.jsonl	25,032 (80%)
validation	validation.jsonl	3,129 (10%)
test	test.jsonl	3,129 (10%)

Total: 31,290 examples

Changes from v1

• Added "karen" class (label 0) for hostile/entitled customers
• Expanded from 5 to 6 sentiment classes
• Labels shifted: karen=0, very_negative=1, negative=2, neutral=3, positive=4, very_positive=5
• Added validation split
• Increased dataset size (~31K vs ~4K)

Intended Use

Demo, experimentation, and educational purposes. No real user data; all samples are synthetic.