Datasets:
Tasks:
Text Classification
Modalities:
Text
Formats:
csv
Sub-tasks:
multi-class-classification
Size:
< 1K
License:
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,25 +1,45 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
---
|
|
|
|
| 4 |
# Trilingual Fraud & Consumer Protection Dataset (V2)
|
| 5 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
## Overview
|
| 7 |
-
This dataset focuses on real-world fraud detection and consumer protection scenarios across English, Hindi, and Punjabi. It is designed to reflect how scams actually appear in everyday communication, often subtle, conversational, and not always obviously malicious.
|
| 8 |
|
| 9 |
-
|
| 10 |
|
| 11 |
-
|
|
|
|
|
|
|
| 12 |
|
| 13 |
## Key Features
|
|
|
|
| 14 |
- Multilingual coverage (English, Hindi, Punjabi)
|
| 15 |
- Code-mixed conversational data common in Indian contexts
|
| 16 |
- Multi-step fraud sequences (trust-building → payment → urgency)
|
| 17 |
- Realistic safe responses (not just advisory, but conversational)
|
| 18 |
- Edge cases that capture misleading or borderline scenarios
|
| 19 |
|
| 20 |
-
---
|
| 21 |
-
|
| 22 |
## Structure
|
|
|
|
| 23 |
Each entry includes:
|
| 24 |
- **text** – message content
|
| 25 |
- **label** – scam or safe
|
|
@@ -29,16 +49,17 @@ Each entry includes:
|
|
| 29 |
- **reasoning_tag** – underlying behavioral signal
|
| 30 |
- **difficulty** – relative complexity
|
| 31 |
|
| 32 |
-
---
|
| 33 |
-
|
| 34 |
## Use Cases
|
|
|
|
| 35 |
- Fraud and scam detection systems
|
| 36 |
- LLM safety and alignment evaluation
|
| 37 |
- Multilingual NLP benchmarking
|
| 38 |
- Consumer protection and risk modeling
|
| 39 |
|
| 40 |
-
---
|
| 41 |
-
|
| 42 |
## Notes
|
| 43 |
-
|
| 44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
---
|
| 4 |
+
|
| 5 |
# Trilingual Fraud & Consumer Protection Dataset (V2)
|
| 6 |
|
| 7 |
+
## Why this dataset matters
|
| 8 |
+
|
| 9 |
+
Fraud detection in regions like North India is not straightforward.
|
| 10 |
+
|
| 11 |
+
Messages are rarely written in clean English. Instead, they are often a mix of Punjabi, Hindi, and English, and many scam messages look very similar to legitimate communication.
|
| 12 |
+
|
| 13 |
+
In practice, the harder part is not spotting obvious scams, but telling apart messages that look almost the same but mean very different things.
|
| 14 |
+
|
| 15 |
+
Many existing datasets focus on clear scam patterns or keyword-based signals. However, this creates a gap where models learn surface patterns instead of actual intent.
|
| 16 |
+
|
| 17 |
+
I built this dataset to try to address that gap by focusing on:
|
| 18 |
+
- contrastive examples (similar wording, different intent)
|
| 19 |
+
- realistic safe responses (not just warnings, but natural conversation)
|
| 20 |
+
- multi-step scam flows where intent evolves over time
|
| 21 |
+
- code-mixed language that reflects how people actually communicate
|
| 22 |
+
|
| 23 |
+
In practice, this turns fraud detection from a pattern recognition problem into an intent understanding problem.
|
| 24 |
+
|
| 25 |
## Overview
|
|
|
|
| 26 |
|
| 27 |
+
This dataset focuses on capturing subtle differences in intent between scam and legitimate messages in real-world multilingual communication across English, Hindi, and Punjabi.
|
| 28 |
|
| 29 |
+
It tries to reflect how scams actually appear in everyday communication, often subtle, conversational, and not always obviously malicious.
|
| 30 |
+
|
| 31 |
+
Instead of only using clear-cut examples, the dataset includes multi-step interactions, edge cases, and realistic safe responses to better capture how intent evolves in real situations.
|
| 32 |
|
| 33 |
## Key Features
|
| 34 |
+
|
| 35 |
- Multilingual coverage (English, Hindi, Punjabi)
|
| 36 |
- Code-mixed conversational data common in Indian contexts
|
| 37 |
- Multi-step fraud sequences (trust-building → payment → urgency)
|
| 38 |
- Realistic safe responses (not just advisory, but conversational)
|
| 39 |
- Edge cases that capture misleading or borderline scenarios
|
| 40 |
|
|
|
|
|
|
|
| 41 |
## Structure
|
| 42 |
+
|
| 43 |
Each entry includes:
|
| 44 |
- **text** – message content
|
| 45 |
- **label** – scam or safe
|
|
|
|
| 49 |
- **reasoning_tag** – underlying behavioral signal
|
| 50 |
- **difficulty** – relative complexity
|
| 51 |
|
|
|
|
|
|
|
| 52 |
## Use Cases
|
| 53 |
+
|
| 54 |
- Fraud and scam detection systems
|
| 55 |
- LLM safety and alignment evaluation
|
| 56 |
- Multilingual NLP benchmarking
|
| 57 |
- Consumer protection and risk modeling
|
| 58 |
|
|
|
|
|
|
|
| 59 |
## Notes
|
| 60 |
+
|
| 61 |
+
The focus of this dataset is on structure, realism, and variation, rather than synthetic or templated generation.
|
| 62 |
+
|
| 63 |
+
Small differences in wording can significantly change intent, and this dataset is designed to capture those nuances.
|
| 64 |
+
|
| 65 |
+
This is still an evolving dataset, and I’m continuing to improve edge cases where even humans might hesitate.
|