Build Jev-Style Decision Models from Scratch with PyTorch and Hugging Face

Learn how to build Jev-style decision-making models from scratch using PyTorch and Hugging Face, replacing token-by-token decoding with a single forward pass to yield predictable outputs such as choices, ordinal scores, and boolean values. The session covers inference principles from the ground up and works toward the RLCD training workflow, using pre-trained models as a foundation.

■ Jev Fundamentals
- Choice, score, and null interfaces, and how a single forward pass differs from traditional decoding
- Traditional LLM inference walkthrough: prefill, decode, and step-by-step JSON generation
■ Implementations
- Encoder-only approach: question-first encoding, mask tokens, and query-option interaction
- Decoder-only approach: causal masking, KV cache reuse, and packing multiple questions
- Encoder vs. decoder trade-offs and guidance on when to choose each
■ Training with RLCD
- Loss design combining cross-entropy with RL components, GRPO-based advantage computation, and training options

Aimed at viewers already comfortable with attention, KV cache, logits, softmax, and PyTorch, the session is inference-focused and does not require a GPU. It provides an intuition for how these models work and a starting template you can adapt to your own use case.