shisha-07/Llama-3-Indian-Gender-Classifier

🇮🇳 Llama-3 Indian Name Gender Classifier

📖 Model Description

This is a fine-tuned version of the Llama-3-8B model, specifically optimized to classify the gender of Indian Names into three categories: Male, Female, and Neutral.

The model was trained using Unsloth for efficient 4-bit fine-tuning and has been exported in multiple formats to support both production (Python/GPU) and local (CPU/Laptop) environments.

🛠️ Source Code

The training code and notebooks are available on GitHub:

• Training Notebook

Key Features

• High Precision for Females: The model achieves 100% precision on female names in validation tests.
• Regional Awareness: Trained specifically on Indian linguistic patterns.
• Multi-Format: Available as LoRA adapters, Merged 16-bit weights, and GGUF.

---

📂 Repository Structure

This repository is organized into three specific folders to keep files clean. Please select the folder that matches your use case:

Folder	Format	Use Case
/Merged	SafeTensors (16-bit)	Best for Python, vLLM, and Production APIs. This is the standalone model with weights baked in.
/GGUF	GGUF (q4_k_m)	Best for Ollama, LM Studio, and Local CPU Inference.
/LoRA	LoRA Adapters	Best for Research, Unsloth, and PEFT loading. Contains only the fine-tuned difference.

---

📊 Evaluation & Performance

The model was evaluated on a held-out test set of 59 diverse Indian names, including difficult unisex/neutral examples.

Summary Metrics

Metric	Score	Interpretation
Accuracy	86.44%	Overall correctness across all categories.
F1-Score	0.8355	Weighted score balancing precision and recall.

Detailed Breakdown

Category	Precision	Recall	F1-Score	Analysis
Male	0.76	1.00	0.87	Catches all males, but sometimes over-predicts male for neutral names.
Female	1.00	0.96	0.98	Extremely reliable. If it says "Female", it is almost certainly correct.
Neutral	1.00	0.22	0.36	Struggles with ambiguity; often defaults to "Male".

⚠️ Known Limitations & Bias

• Male Defaulting: The model has a strong bias to classify ambiguous or neutral names as Male.
  • Failed Examples: Kiran, Sonu, Suman, Gurpreet (Predicted: Male).
• Specific Mismatches: The name Simran was incorrectly predicted as Male in our validation set.
• Scope: This model is strictly for Indian names and may not perform accurately on Western or East Asian names.

---

🚀 How to Use

1. Python (Transformers) - Recommended

Use this method if you are running a script or a web server. Note: You must specify subfolder="Merged".

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# 1. Load from the 'Merged' folder
model_id = "shisha-07/Llama-3-Indian-Gender-Classifier"
tokenizer = AutoTokenizer.from_pretrained(model_id, subfolder="Merged")
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    subfolder="Merged",
    device_map="auto",
    torch_dtype=torch.float16
)

# 2. Define the Prompt (Alpaca Format)
prompt_template = """### Instruction:
Identify the gender of the given Indian name.

### Input:
{}

### Response:
"""

# 3. Run Inference
name = "Ananya"
inputs = tokenizer(prompt_template.format(name), return_tensors="pt").to("cuda")
output = model.generate(**inputs, max_new_tokens=10)

print(tokenizer.decode(output[0], skip_special_tokens=True))
# Output: Female

2. GGUF (Ollama / Local) -

Download the .gguf file from the GGUF Folder.

• CLI Command:

./llama-cli -m llama-3-indian-gender.Q4_K_M.gguf -p "### Instruction: Identify the gender. ### Input: Rajesh ### Response:"

• Ollama: Create a Modelfile with the following content:

FROM ./llama-3-indian-gender.Q4_K_M.gguf
TEMPLATE """### Instruction:
Identify the gender of the given Indian name.

### Input:
{{ .Prompt }}

### Response:
"""

3. LoRA Adapters (Unsloth) -

Use this if you want to load the adapters on top of the base Llama-3 model dynamically.

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "shisha-07/Llama-3-Indian-Gender-Classifier",
    subfolder = "LoRA", # Specific folder for adapters
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# ... Run inference as usual

---

🛠️ Training Details

• Base Architecture: Llama-3-8B-bnb-4bit

• Fine-Tuning Method: QLoRA (via Unsloth)

• Dataset: Custom dataset of Indian names labeled by gender.

• Infrastructure: Trained on Tesla T4 GPU (Google Colab)

📜 License

Apache-2.0