LLM News Classifier Interface

A hybrid news article classification system that combines TF-IDF shortlisting with fine-tuned Mistral-7B-LoRA inference. Articles are classified into 17 categories using a two-stage pipeline that balances speed and accuracy, served via a Django REST API with an interactive web UI.

How It Works

1. TF-IDF Shortlist — Word and character n-gram vectorizers rank the top-N most likely categories using cosine similarity against pre-trained label centroids.
2. LLM Inference — A 4-bit quantized Mistral-7B-Instruct model (LoRA-adapted) performs restricted decoding, scoring only the shortlisted candidate tokens. This prevents hallucinations and keeps inference fast.
3. Chunk Fusion — Long articles are split into overlapping windows. Each window is scored independently, then logits are fused across chunks using a configurable pooling strategy.

Categories

Letter	Category	Letter	Category
A	arts	J	lifestyle
B	crime	K	politics
C	disaster	L	religion
D	economy	M	science
E	education	N	social
F	environmental	O	sport
G	health	P	unrest
H	humanInterest	Q	weather
I	labour

Tech Stack

• Backend: Django 6, Django REST Framework
• ML: PyTorch, Hugging Face Transformers, PEFT (LoRA), scikit-learn, SciPy
• Quantization: BitsAndBytes (4-bit NF4, double quant)
• Base Model: mistralai/Mistral-7B-Instruct-v0.3
• Frontend: Vanilla HTML/CSS/JavaScript

Project Structure

llm_classifier_interface/
├── llm_classifier_interface/
│   ├── interface.py          # Core ML pipeline (TF-IDF + LLM)
│   ├── views.py              # Django endpoints
│   ├── urls.py               # URL routing
│   ├── settings.py           # Django config
│   ├── tfidf_assets/         # Vectorizers, centroids, meta.json
│   └── mistral_lora_06012026/# LoRA adapter weights + tokenizer
├── templates/
│   └── index.html            # Web UI
└── manage.py

Setup

Requirements

pip install django djangorestframework torch transformers peft bitsandbytes scikit-learn scipy joblib

Note: BitsAndBytes 4-bit quantization requires a CUDA-capable GPU with at least ~6GB VRAM.

Model Assets

Place the following in the expected directories (or override via environment variables):

• llm_classifier_interface/mistral_lora_06012026/ — LoRA adapter weights and tokenizer config
• llm_classifier_interface/tfidf_assets/ — tfidf_word.joblib, tfidf_char.joblib, centroids/, meta.json

Override paths with environment variables:

export LORA_DIR=/path/to/lora_adapter
export TFIDF_DIR=/path/to/tfidf_assets

Run

python manage.py migrate
python manage.py runserver

The model loads on the first request and is reused for all subsequent requests.

Usage

Web UI

Open http://localhost:8000/ in a browser. Paste an article, adjust parameters, and click Classify.

REST API

POST /api/predict/
Content-Type: application/x-www-form-urlencoded

Parameter	Type	Default	Description
text	string	required	Article text to classify
topN	int	5	Number of TF-IDF candidate labels to shortlist
k_chunks	int	4	Number of top-scoring chunks to run through LLM
mode	string	hybrid	Logit fusion mode: max, sum, or hybrid
alpha	float	0.7	Weight of sum in hybrid mode (sum·α + max·(1-α))

Example response:

{
  "label": "politics",
  "confidence": 0.8731,
  "candidates": [
    {"label": "politics", "score": 0.412},
    {"label": "economy",  "score": 0.289},
    {"label": "social",   "score": 0.201}
  ]
}

Architecture Notes

• Restricted decoding: The model only scores the 17 valid category tokens — output is always a valid label.
• Chunk ranking: Windows are ranked by TF-IDF relevance before LLM inference; only the top k_chunks are processed, keeping latency bounded regardless of article length.
• Singleton model: NewsClassifierService is instantiated once per process and shared across all requests.
• Hybrid fusion: mode=hybrid blends sum pooling (rewards consistent agreement across chunks) with max pooling (rewards a single strong signal), controlled by alpha.