It distinguishes between genuine anti-India propaganda (misinformation, manipulated narratives) and legitimate criticism, balanced reporting, or academic analysis.

📋 Propaganda Categories Detected

Category	Description
Territorial Integrity	False claims about Kashmir, Arunachal Pradesh, borders
Religious Division	Content inciting communal hatred and exaggerating tensions
Economic Misinformation	Distorted facts about India's economy and development
Historical Revisionism	Distorting Indian history and civilization
Geopolitical Undermining	Narratives to weaken India's global standing
Military/Security FUD	Exaggerated threat narratives, false military claims
Separatist Propaganda	Content promoting balkanization of India
Democratic Institution Attacks	Undermining trust in democracy, judiciary, elections

🔧 Training Details

Parameter	Value
Base Model	Qwen/Qwen2.5-0.5B
Method	QLoRA (4-bit NF4 + LoRA)
LoRA Rank	16
LoRA Alpha	32
Target Modules	q_proj, k_proj, v_proj, o_proj
Training Samples	~86 curated examples
Epochs	5
Learning Rate	2e-4 (cosine schedule)
Optimizer	paged_adamw_8bit
Max Seq Length	512

💻 Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

# Load model
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B",
    quantization_config=bnb_config,
    device_map="auto",
)

model = PeftModel.from_pretrained(base_model, "ProfRutPatel/Qwen2.5-0.5B-India-Propaganda-Detector")
tokenizer = AutoTokenizer.from_pretrained("ProfRutPatel/Qwen2.5-0.5B-India-Propaganda-Detector")

# Analyze text
system_prompt = "You are an expert analyst specialized in detecting anti-India propaganda..."
text = "Your text to analyze here"

prompt = f"<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\nAnalyze the following text for anti-India propaganda:\n\n\"{text}\"<|im_end|>\n<|im_start|>assistant\n"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.1)
result = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(result)

⚠️ Limitations

Trained on a curated dataset of ~86 examples; may not cover all propaganda patterns
Best used as a screening tool, not a definitive classifier
Should be combined with human review for critical applications
May not perform well on very subtle or novel propaganda techniques

🏗️ Hardware

Trained on NVIDIA GeForce GTX 1650 (4GB VRAM)
QLoRA enables training on consumer GPUs