📚 Anki Card Generator with LoRA-Finetuned LLM

Overview

This project aims to fine-tune a Large Language Model (LLM) using LoRA to automatically generate high-quality Anki flashcards from various knowledge sources (Wikipedia or Personal pdfs).

The system supports:

• Automatic training data generation from PDFs and Wikipedia pages
• LoRA fine-tuning
• Flashcard generation from PDFs
• Flashcard generation from Wikipedia pages

The goal is to produce concise, factual, and pedagogically effective Anki cards suitable for long-term learning.

Models used fro data generation and lora fine-tuning can be ajusted to the user's computing power.

My personal fine-tuned model is available on this link: https://huggingface.co/Guibibo/Mistral-7B-v0.3-FlashCards.

---

Features

• LoRA fine-tuning
• Domain-agnostic flashcard generation
• PDF parsing and knowledge extraction
• Wikipedia page ingestion
• Synthetic training data generation
• Anki-compatible outputs (APKG)
• Modular end-to-end pipeline

---

Usage

Installation

pip install -r environment.yml

Data and cards generation

python3 -m main

JSON datasets combination

python3 -m data.combine_data data/revised/file1.jsonl data/revised/file2.jsonl ...

JSON datasets adaptation of generated data from LLama model to LoRA on a Mistral model

python3 -m data.convert_to_lora.py path/to/file

LoRA fine-tunning

python3 -m LoRA.run_lora