everettjf/ezjev-4b-s3
The everettjf/ezjev-4b-s3 is a 4.5 billion parameter typed-decision model, built upon the Qwen3.5-4B architecture and fine-tuned through multiple stages. It is specifically designed to answer 'choice', 'noul', and 'score' questions with a probability distribution per option in a single forward pass, without generating tokens. This model excels at complex programmatic decision-making tasks, particularly those involving multi-clause policies, deadlines, and rule lookups, making it suitable for business document analysis and automated reasoning.
Loading preview...
ezjev-4b-s3: A Specialized Typed-Decision Model
The ezjev-4b-s3 is a 4.5 billion parameter model developed by everettjf, specifically engineered for typed-decision tasks. Unlike traditional generative LLMs, it provides probabilistic answers to structured questions (choice, noul, score) in a single forward pass, eliminating token generation. This model is built on the Qwen3.5-4B lineage, refined through a multi-stage LoRA fine-tuning process.
Key Capabilities
- Typed-Decision Making: Directly outputs probabilities for predefined options, ideal for structured queries.
- Complex Rule Interpretation: Excels at processing intricate business logic, including multi-clause policies, time-based deadlines, unit conversions, and alias lookups.
- Programmatically Generated Training: Stage 3 training data consists of ~20k programmatically generated decisions, simulating long business documents with random scenarios and gold labels computed by code, ensuring robust reasoning capabilities.
- Efficient Inference: Designed for efficient serving with vLLM, requiring approximately 9 GB of GPU memory.
Good for
- Automated Business Logic: Ideal for applications requiring precise, automated decision-making based on complex rules and policies.
- Document Analysis: Suitable for extracting structured decisions from legal, financial, or operational documents.
- Reasoning Tasks: Particularly strong in scenarios demanding logical inference and handling under-specified cases with a "cannot determine" label.
- High-Throughput Decision Systems: Its single-pass, non-generative nature makes it efficient for systems needing rapid, probabilistic answers to structured questions.