systemonedev/kenning-xl-v0.6
systemonedev/kenning-xl-v0.6 is an experimental 2-billion parameter decision model built on a Qwen3-1.7B decoder, designed to answer typed questions with calibrated probabilities. It features a constrained answer-token readout and an optional deliberate reasoning mode for complex decisions. This model excels particularly in agent and structured-record decision tasks, offering enhanced reasoning capabilities over traditional cross-encoders.
Loading preview...
Kenning-XL v0.6: A Decoder-Based Decision Model
Kenning-XL v0.6, developed by System One, is an experimental 2-billion parameter decision model built upon a Qwen3-1.7B decoder. Unlike traditional cross-encoders, it processes typed questions to produce calibrated probabilities by allowing the full language model to reason over the entire state in a single pass. A key feature is its constrained answer-token readout and an optional deliberate mode that generates a short reasoning trace for numeric, policy, and table decisions where a single pass is insufficient.
Key Capabilities & Features
- Decoder-based Architecture: Utilizes a Qwen3-1.7B base model, fine-tuned with LoRA.
- Constrained Answer Readout: Renders (state, question) and reads the next-token distribution over answer tokens in one pass.
- Deliberate Reasoning Mode: Activates for low-confidence or structured questions, generating a reasoning trace before answering, improving accuracy on complex tasks like arithmetic.
- Specialized Training: Incorporates readout distillation on a structured corpus and gold-trace deliberate training, where the model learns from exact arithmetic/threshold/membership reasoning chains.
Performance Highlights
Benchmarking against kenning v0.5, Clef, and Jev, Kenning-XL v0.6 demonstrates strong performance:
- Agent Decisions: Achieves 0.860, significantly outperforming
kenning v0.5(0.727) andClef(0.793). - Structured Records & Tables: Shows improved performance over
kenning v0.5(0.663 vs 0.587 for Records; 0.660 vs 0.520 for Tables). - Arithmetic Tasks: Deliberate reasoning more than doubles performance on hard arithmetic tasks (e.g., daily-limit checks 0.42 → 0.81).
When to Use This Model
Kenning-XL v0.6 is particularly strongest on agent and structured-record decisions. It is recommended for use cases involving:
- Agent tool calls and task completion.
- Rules over JSON records and tables.
- Complex decisions requiring explicit reasoning traces.
For tasks involving logs, quality assessment, or general text, the kenning v0.5 cross-encoder might still be more suitable as Kenning-XL currently trails in these areas.