How to train INSANELY fast JEV Decision Models from scratch (architecture deep-dive)

An implementation of SystemOne Decision Models inspired by Jev. Check Prime Intellect here: https://www.primeintellect.ai/?utm_source=youtube&utm_medium=video&utm_campaign=sv001&utm_content=avb_fj Follow me on Twitter: https://x.com/neural_avb Support on Patreon: https://www.patreon.com/NeuralBreakdownwithAVB Members get access to everything behind-the-scenes that goes into producing my videos - including code. Plus, it supports the channel in a big way and helps to pay my bills. Training repo: https://github.com/avbiswas/bev-train HF Dataset: https://huggingface.co/datasets/avbiswas/bev-decision Model: https://huggingface.co/avbiswas/bev-decider-0.4B Livesteam coding: https://youtube.com/live/AzxoU7kxjig Timestamp: 0:00 Intro: System One decision models 1:38: What is Jev? 3:43 What we're building 8:08 Choice, score & noul questions 12:25 Bi-encoders vs joint encoders 14:26 Designing the network on Qwen 17:55 The decision head 22:49 Making it choice-order invariant 32:39 Live coding the model in PyTorch 1:31:46 Building the dataset 1:41:35 Training & results