🔒Jev: A Fast, Local LLM for Text Classification
A new local LLM for text classification without cloud dependency
TL;DR
Jev, a local large language model, classifies text into Legitimate, Spam, and Phishing categories without sending data to external servers. It's trained on synthetic data and uses Reinforcement Learning for Calibrated Decisions (RLCD).
Jev, a local large language model, is making waves for its ability to classify text into Legitimate, Spam, and Phishing categories without sending data to external servers. If you're working on security or spam detection, this could be a game-changer. Jev is trained on synthetic data and uses Reinforcement Learning for Calibrated Decisions (RLCD) to improve decision accuracy. The model outputs probabilities with a precision of three decimal places, making it highly reliable for automated decision-making systems. Jev is open-source and available on GitHub, offering developers a powerful tool for local text classification tasks.
Key Points
Jev is a large language model trained on synthetic data using the huggingface-hub library.
The model requires Python 3.12 or higher and depends on huggingface-hub, llama-cpp-python, and numpy.
Jev is loaded using Llama.from_pretrained with a context length of 512 and logits_all=True.
The model outputs probabilities for Legitimate (0.031), Spam (0.084), and Phishing (0.885) categories.
Jev is part of the NobodyWho project and was published by Duarte O.Carmo Technical on September 22, 2026.
Why It Matters
If you're working on a security system that needs to classify text locally, Jev's ability to run without external server dependencies could save you a lot of bandwidth and latency. The model's use of synthetic data and RLCD also means it's highly accurate, even in environments with limited data.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,487 builders reading daily.