CJRouter/qwen3.5-9b-clef-decision
CJRouter/qwen3.5-9b-clef-decision is a 9 billion parameter language model based on Qwen3.5-9B, specifically fine-tuned with a Cloudflare Clef-style `joint_head` decision architecture. This model is designed to produce structured, calibrated-probability decisions for complex business scenarios, offering choice probability distributions and confidence scores. Its primary use case is structured decision-making in applications like agent trace observability, customer service ticketing, invoice processing, and security incident response, providing fine-grained, per-question accuracy and calibrated confidence.
Loading preview...
Qwen3.5-9B-Clef Decision Model Overview
This model is a specialized 9 billion parameter language model built upon the Qwen3.5-9B base, featuring a unique Cloudflare Clef-style joint_head decision architecture. It is engineered to deliver structured, calibrated-probability decisions for intricate business contexts, providing choice probability distributions and confidence levels for each decision point.
Key Capabilities
- Structured Decision-Making: Outputs precise choice probability distributions and confidence scores for defined questions based on state descriptions.
- Calibrated Confidence: Achieves an Expected Calibration Error (ECE) of 0.0355 on an independent test set, ensuring that stated confidence levels accurately reflect the probability of correctness. This is crucial for automated risk-threshold decisions.
- Fine-grained Per-Question Accuracy: Evaluated with per-question accuracy (73-84% across domains), allowing downstream systems to set individual confidence thresholds for each decision item.
- Rapid Training: The model was fine-tuned in approximately 21 minutes on a single NVIDIA H20 GPU using LoRA and joint decision-head training, demonstrating efficient resource utilization.
- Private and Sovereign: Licensed under Apache-2.0, designed for on-premise deployment without vendor lock-in or data egress, ensuring sensitive business data remains within infrastructure.
Good For
- Agent Trace Observability: Judging autonomous agent run traces for continuation, human intervention needs, task outcomes, and risk levels.
- Customer Service Ticketing: Automatically determining ticket urgency, assigning teams, and assessing customer sentiment.
- Invoice Processing: Making structured judgments on various invoice-related tasks.
- Security Incident Response: Assessing incident risk levels and prioritizing handling actions.
- Applications requiring defensible, automated risk-threshold decisions based on reliable confidence scores.