karanverma19/tri-fraud-alerts-india

⚠️ This dataset has been updated. Latest version: https://huggingface.co/datasets/karanverma19/trilingualfraudconsumerprotectionv2

Trilingual (Punjabi–Hindi–English) Fraud & Consumer Protection Dataset

Why I built this

Fraud detection in regions like Punjab is not straightforward. Messages are rarely in clean English and are often written in Punjabi, Hindi, or a mix of both. On top of that, many scam messages look very similar to legitimate communication.

While working on this dataset, I noticed that most existing datasets:

• —focus heavily on obvious scam examples
• —don’t include enough real “safe” messages
• —ignore how people actually communicate (code-mixed language)

This dataset is an attempt to address those gaps.

What this dataset is trying to do

Instead of just collecting examples, I tried to structure the dataset in a way that helps models make better decisions.

Contrastive examples

For many cases, I included pairs of messages that look similar but have different intent. The idea is to help a model learn why something is a scam, not just what looks like a scam.

Negative space

A portion of the dataset contains normal, safe communication. This is important because, in real systems, over-flagging is just as problematic as missing scams.

Edge cases

Some messages are intentionally borderline. These are cases where the language is subtle and not obviously fraudulent, which is closer to real-world scenarios.

Code-mixed language

A lot of the data includes Punjabi–Hindi–English mixing. This reflects how people actually speak and message in North India.

Languages

The dataset includes:

• —Punjabi
• —Hindi
• —English

Many samples are code-mixed across these languages.

What changed during development

The initial versions of the dataset had a few problems:

• —patterns were repetitive
• —scam signals were too obvious
• —safe examples were limited

To improve this:

• —I added contrastive pairs across categories
• —included neutral and advisory messages
• —introduced reasoning tags like urgency, authority, and guaranteed outcomes
• —improved Punjabi and Hinglish phrasing to feel more natural

Some trade-offs remain, especially between keeping messages realistic and maintaining clear labeling boundaries.

Dataset structure

• —text
• —label (scam / safe)
• —category (fraud / consumer_protection)
• —sub_type (visa, job, otp, etc.)
• —language
• —code_mixed
• —contexttype (original / contrast / negative / edgecase)
• —reasoning_tag
• —difficulty

Key observation

In this dataset, many fraud messages rely on:

• —urgency (“act now”)
• —authority claims (“embassy connection”)
• —guaranteed outcomes (“100% visa”)

These signals are often embedded in code-mixed language, which makes them harder to detect.

Limitations

• —The dataset is synthetic, not collected from real user logs
• —It may not capture all regional dialect variations
• —It is focused mainly on immigration and related fraud patterns

Ethical note

All samples are synthetically generated based on patterns observed in publicly reported fraud cases. No personal or sensitive data is included.

What changed in this iteration

I didn’t try to increase dataset size in this version. The focus was more on improving how clearly a model can distinguish between scam and safe messages.

Some changes I made:

• —added more subtle edge cases where intent is not obvious
• —expanded reasoning patterns like social proof and fake authority
• —improved Punjabi–Hindi code-mixed phrasing to feel more natural
• —made contrast between scam and safe messages clearer

After running it through Adaptive Data:

Grade improved from E → A Quality score improved from 2.0 → 9.8 (~390% improvement) Percentile: 0.1 → 57.7

One thing that stood out during this process is that structure matters more than size. Even small changes in phrasing or context can shift how a message is interpreted.

Acknowledgment

This dataset was developed and refined using the Adaptive Data platform by Adaption Labs as part of the Uncharted Data Challenge.

Adaptive Data was used to improve phrasing consistency, align scam and safe examples, and strengthen intent clarity.

Platform: Adaption Labs – Adaptive Data