everettjf/ezjev-4b-s2
The everettjf/ezjev-4b-s2 is a 4.5 billion parameter typed-decision model, fine-tuned from Qwen/Qwen3.5-4B, designed to answer 'choice', 'noul', and 'score' questions with a probability per option in a single forward pass. It excels in decision-making tasks, achieving the best performance among models ≤ 5B parameters on the Jev Decision Index 0.3 leaderboard. This model is optimized for tasks requiring nuanced decision outputs, such as RAG hallucination detection, product relevance, and code-output selection.
Loading preview...
ezjev-4b-s2: A Specialized Typed-Decision Model
ezjev-4b-s2 is a 4.5 billion parameter language model, fine-tuned from Qwen/Qwen3.5-4B, specifically designed for typed-decision tasks. It processes 'choice', 'noul', and 'score' questions, providing a probability for each option in a single forward pass. This model is a merged LoRA fine-tune, developed in two stages to enhance its decision-making capabilities across various domains.
Key Capabilities
- Typed-Decision Making: Optimized to output structured decisions (choice, noul, score) with associated probabilities.
- Strong Performance in Decision Index: Achieved a Full score of 46.95 on the official Jev Decision Index 0.3, ranking as the best performer among all models ≤ 5B parameters.
- Targeted Fine-tuning: Stage 2 fine-tuning focused on improving performance in weak task families, including RAG hallucination, product relevance, code-output selection, and claim verification.
- Efficient Inference: Demonstrated a median latency of 35.5 ms for 150,759 requests on one RTX PRO 6000 during Decision Index 0.2.1 evaluation.
Good For
- Automated Decision Systems: Ideal for applications requiring probabilistic decision outputs from text.
- Benchmarking against Jev Decision Index: A strong candidate for tasks evaluated by the Jev Decision Index, particularly for its size class.
- Specialized NLP Tasks: Suitable for use cases involving RAG hallucination detection, product relevance scoring, and structured question answering where a 'typed-decision' output is beneficial.
For more details, including quickstart guides and downloads, visit the project page.