Prompt48/my-jev-4b
Prompt48/my-jev-4b is a 4.5 billion parameter Jev-like System One classifier, fine-tuned from Qwen3.5-4B with Unsloth. This model is specifically designed to classify input `state` and `question` against a list of `options`, returning a single letter. It excels at decision tasks, achieving 87.2% accuracy on a held-out development set, making it suitable for intent recognition and policy routing.
Loading preview...
my-jev-4b: A Specialized System One Classifier
Prompt48/my-jev-4b is a 4.5 billion parameter model, fine-tuned from Qwen3.5-4B using Unsloth. Its primary function is to act as a Jev-like System One classifier, designed to process a given state, question, and a list of options, then output a single letter representing the best match. This model is optimized for precise classification tasks.
Key Capabilities
- High Accuracy Classification: Achieves 87.2% accuracy on a held-out development set (594 examples), significantly outperforming the base Qwen3.5-4B model (72.6%).
- Specialized Task Handling: Specifically trained to evaluate decision tasks and select one option, returning only its letter.
- Improved Performance Across Domains: Demonstrates notable accuracy improvements across various sources, including
policy_v2(from 53.0% to 95.5%) androuting_v2(from 54.5% to 87.9%). - Efficient Fine-tuning: Fine-tuned using Unsloth on the open tev1 "new-v1" dataset, comprising 37,840 examples.
Good For
- Intent Recognition: Classifying customer messages or queries into predefined support intents.
- Policy Routing: Directing requests or decisions based on specific policies.
- Structured Decision Tasks: Any application requiring a model to select a single, definitive option from a given set based on context.
- System One Classification: Tasks where a quick, confident, and single-choice output is required, without explanation.