kerala-crime-detective-gemma

🔍 Kerala Crime Detective — Malayalam + English + Manglish AI

<p align="center"> <img src="https://img.shields.io/badge/Language-Malayalam%20%7C%20English%20%7C%20Manglish-blue" /> <img src="https://img.shields.io/badge/Base%20Model-Gemma%203%201B-orange" /> <img src="https://img.shields.io/badge/Task-Crime%20Solving%20%2B%20Comedy-green" /> <img src="https://img.shields.io/badge/License-Apache%202.0-red" /> </p>

Solve crimes the Kerala way — Comedy, Manglish, and serious detective work, all in one model!

A fine-tuned Gemma 3 1B that understands Kerala crime reports, FIR details, and cyber fraud cases — and responds in Malayalam, English, or Manglish with comedy and serious investigation steps.

---

🎯 What This Model Does

---

🚀 Try the Live Demo

👉 Open in HuggingFace Spaces

---

📦 Model Details

---

💻 Quick Start

Installation

pip install transformers accelerate torch

Basic Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_ID = "wincode/kerala-crime-detective-gemma"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="eager"
)
model.eval()

def solve_crime(crime_report: str, mode: str = "comedy") -> str:
    system_prompts = {
        "comedy": (
            "You are Kerala Nadodikattu Detective, a comedy crime solver. "
            "Solve crimes using Malayalam, Manglish and English mix. "
            "Use humor, local references, Kerala culture."
        ),
        "serious": (
            "You are a senior Kerala Police CID officer. "
            "Analyze crime reports professionally with evidence analysis, "
            "suspect profiling, investigation steps and legal sections."
        ),
        "cyber": (
            "You are Kerala Cyber Cell's top investigator. "
            "Specialize in UPI fraud, SIM swap, sextortion, fake jobs. "
            "Provide immediate victim steps, recovery options, legal recourse."
        ),
    }

    messages = [
        {"role": "system", "content": system_prompts.get(mode, system_prompts["comedy"])},
        {"role": "user",   "content": crime_report},
    ]

    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
    )

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=400,
            do_sample=True,
            temperature=0.8,
            top_p=0.9,
            repetition_penalty=1.1,
            pad_token_id=tokenizer.pad_token_id,
        )

    new_tokens = output_ids[0][inputs["input_ids"].shape[1]:]
    return tokenizer.decode(new_tokens, skip_special_tokens=True)


# ── Example 1: Malayalam Comedy Mode ─────────────────────────
report1 = """
Crime Report:
Location: Thrissur Pooram grounds
Time: 11 PM
Crime: 2kg gold ornaments missing from elephant caparison
Evidence: Footprints, torn dhoti piece
FIR: THR/2024/445
"""
print(solve_crime(report1, mode="comedy"))


# ── Example 2: Cyber Crime Mode ───────────────────────────────
report2 = """
Cyber Crime FIR:
Victim: Anitha, Homemaker, Palakkad
Crime: Received WhatsApp message saying I won Rs 50 lakh lottery.
They asked Rs 15,000 processing fee. I paid. Number now switched off.
Amount lost: Rs 15,000
"""
print(solve_crime(report2, mode="cyber"))


# ── Example 3: Serious Investigation Mode ─────────────────────
report3 = """
Hit and Run:
Location: NH-66 Kannur Bypass
Time: 11:30 PM
Victim: Biker, critical, ICU
Evidence: White paint transfer on victim bike
Witness: Trucker saw partial plate KL-13
"""
print(solve_crime(report3, mode="serious"))

---

🗂️ Training Dataset

Trained on the Kerala Crime Comedy Dataset — a custom dataset covering:

Dataset stats:

• Total examples: 21+ (growing)
• Languages: Malayalam, English, Manglish
• Format: Chat (system + user + assistant)
• Average response length: 300-500 tokens

---

📊 Example Outputs

🎭 Malayalam Comedy Mode

Input:

Crime: Gold chain snatching at Calicut beach. Victim: teacher. Suspect: young male on motorcycle.

Output:

⚡ Chain Snatching Investigation Report!

Chechi! Chain poyi aano?
Bike-il vannu, chain eduthu, bike-il poyi!
Classic Calicut evening crime!

🔍 Evidence Analysis:
- Skid marks: Direction analysis possible
- Time: Evening = witnesses available
- Victim is teacher: Respected community member

Serious Steps:
✅ Alert gold shops in 20km radius IMMEDIATELY
✅ Deploy two-wheeler patrol on beach road
✅ CCTV from SM Street main junction
✅ Recovery chance: 60% if acted within 6 hours!

Chechi, we will get justice! 💪

🌐 Cyber Crime Expert Mode

Input:

UPI fraud: Someone sent QR code saying it will send me money. I scanned and entered PIN. Rs 35,000 deducted.

Output:

💸 UPI QR Scam - Classic Trap!

CALL 1930 RIGHT NOW (Cyber Crime Helpline - Free, 24/7)
Every minute counts for account freeze!

How the scam worked:
In UPI there are TWO types of QR codes:
📥 COLLECT request = You are PAYING (needs PIN = money GOES OUT)
📤 PAYMENT = You are RECEIVING (NO PIN needed ever)

GOLDEN RULE: You NEVER enter PIN to RECEIVE money!
If someone asks PIN to send you money = 100% SCAM always!

Recovery steps:
1. Call 1930 immediately
2. Report to your bank fraud line
3. Screenshot the UPI ID and report on cybercrime.gov.in

Recovery probability: 50-60% if reported within 4 hours!

---

⚠️ Limitations

• Model is fine-tuned on a small dataset (21 examples) — responses may not always be perfectly formatted
• Malayalam script quality depends on base model's multilingual capability
• For real emergencies, always contact actual Kerala Police: 100 or Cyber Crime: 1930
• Model provides educational and entertainment value — not a substitute for real legal advice
• Responses may vary due to sampling temperature

---

🛡️ Important Disclaimer

This model is for educational and entertainment purposes only.

For real crimes and emergencies:

• Police Emergency: 100
• Cyber Crime Helpline: 1930
• Women's Helpline: 1091
• Child Helpline: 1098
• Cybercrime Portal: cybercrime.gov.in

---

🏋️ Training Details

# Fine-tuning configuration used
sft_config = SFTConfig(
    max_length=1024,
    num_train_epochs=5,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,   # effective batch = 16
    gradient_checkpointing=True,
    learning_rate=2e-5,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    weight_decay=0.01,
    bf16=True,
    optim="adamw_torch_fused",
)

Hardware used: Kaggle T4 GPU (15GB VRAM) Training time: ~25 minutes for 5 epochs

---

🗺️ Roadmap

• Expand dataset to 500+ examples
• Add more Malayalam script examples
• Add Manglish-only mode
• Support for audio input (voice crime reports)
• Add more cyber crime patterns (2024-2025 new scams)
• Quantized version (GGUF) for local deployment
• API endpoint for police department integration

---

📁 Related Resources

---

🙏 Credits

• Base Model: Google Gemma 3 — Thank you Google DeepMind
• Fine-tuning: HuggingFace TRL — SFTTrainer
• Training Platform: Kaggle — Free T4 GPU
• Demo Framework: Gradio
• Inspiration: Kerala Police, Kerala comedy films, and every aunty who knows everything 🙏

---

📜 License

This model is released under the Apache 2.0 License.

The base model Gemma 3 is subject to Google's Gemma Terms of Use.

---

<p align="center"> Made with ❤️ in Kerala 🌴 | <b>Nammude Kerala, Nammude Detective!</b> </p>