Local LLM Text Classification

This project implements a text classification system powered by Large Language Models (LLMs) running locally. The goal is to leverage the capabilities of modern LLMs to automatically categorize and label text data without relying on external APIs or manual human labeling, ensuring privacy, autonomy, and efficiency in text processing tasks.

• Inference engine: Ollama
• Target models:
  • gpt-oss:20b
  • llama3.1:8b

Project Context: NoNoise LinkedIn ML Feature

The example usecase of this project will be to enable the Implementation of a new feature in the NoNoise LinkedIn browser extension to remove ads posted by people in our network using Machine Learning.

• Upstream project: NoNoise LinkedIn
• Role of this repository: this project is being used to label a dataset of LinkedIn posts, which will then be used to train the Machine Learning model for the above feature.
• Prompt Engineering Technique: this project will use Few-Shot prompting

Why Local LLMs for Classification?

• Privacy by default: data never leaves your machine.
• No external API costs or rate limits.
• Rapid iteration: tweak prompts/labels and re-run instantly.

Prerequisites

• macOS, Linux, or Windows machine with sufficient RAM/VRAM.
• Ollama installed and running:
  • macOS: brew install ollama && ollama serve
  • Linux: curl -fsSL https://ollama.com/install.sh | sh
  • Windows: see https://ollama.com/download
  • Ollama documentation
• Pull the required models:

  ollama pull gpt-oss:20b
  ollama pull llama3.1:8b

Repository Structure

.
├─ csvFiles/
│  └─ linkedin_posts.csv
│  └─ output_data.csv
├─ dataClassification.py
├─ LocalLLMtextClassification.py
├─ promptConstants.py
│
├─ README.md
├─ LICENSE
└─ .gitignore

• csvFiles/linkedin_posts.csv
  • Source dataset of raw LinkedIn post texts (first column is treated as the post body).
  • Replace/extend this file with your own data/use-case to re-run classification.
• csvFiles/output_data.csv
  • Generated output written by dataClassification.py.
• dataClassification.py
  • Batch pipeline entrypoint.
  • Loads csvFiles/linkedin_posts.csv, calls classify_advertisement() for each row using a locally running LLM.
  • logs [AD]/[NOT_AD] (Binary Classification), and writes csvFiles/output_data.csv with a new processed_output column.
• LocalLLMtextClassification.py
  • Thin client for the local Ollama server.
  • Sends the prompt, using fewShotPrompt(text) to the selected model (gpt-oss:20b or llama3.1:8b).
  • Returns the classification label string.
• promptConstants.py
  • Central place for the prompt, labels, and few-shot examples used by the classifier.
  • Tweak this to adjust behavior, labels, or instructions.

Model Selection

• gpt-oss:20b

  • Pros: Higher instruction-following quality, better few-shot adherence.
  • Cons: Heavier, uses more VRAM/RAM, slower on CPU-only.
  • Use when: label space is nuanced or you need stronger reasoning.

• llama3.1:8b

  • Pros: Fast, lightweight, good baseline quality, great for batch throughput.
  • Cons: Can occasionally deviate from strict formatting.
  • Use when: you need speed or are running on limited hardware.

Acknowledgment

• Dataset: 1600 Posts on LinkedIn (Kaggle)
• Ollama for enabling local inference
• Llama (Meta AI), for llama3.1:8b
• OpenAI for gpt-oss:20b model

License

This project is licensed under the GNU General Public License v3.0.