kenning-xl-v0.6 (experimental)
Kenning-XL is a System One decision model built on a decoder. It answers the same wire format as Kenning (POST /v1/systemone; typed questions → calibrated probabilities, no text), but instead of a cross-encoder it uses a small decoder (Qwen3-1.7B) with a constrained answer-token readout: it renders (state, question) and reads the next-token distribution over the answer tokens in one pass, so the full language model reasons over the whole state. An optional deliberate mode generates a short reasoning trace, then reads the answer — for numeric, policy and table decisions a single pass can't do.
This is the experimental v0.6 engine. The published cross-encoder (kenning-large-v0.5) is still the fast default; Kenning-XL is complementary — strongest on agent and structured-record decisions.
How it's served
It is served by SystemOne Builder's kenning-xl engine (systemone_builder.kenning.xl_serve), which answers one-pass and escalates to deliberate mode only for low-confidence, short/structured questions — so the extra latency is spent only where reasoning helps. See the design notes. It is not a plain transformers classifier; load it through the builder.
Training
- Base:
Qwen/Qwen3-1.7B(Apache-2.0), LoRA fine-tuned (merged), bf16, 1,536-token context. - Readout distillation on the v0.5 structured corpus (records, tables, agent steps, logs, text, safety; exact or public labels), plus gold-trace deliberate training: the builder's rule generators compute each label, so they emit the exact arithmetic/threshold/membership reasoning chain for free (
kenning/traces.py), and the model learns to reason-then-answer on those. - Not trained on any TypeSafe output.
Benchmark (general suite, per-family mean, 1,328 held-out items)
As served (confidence + size-gated escalation), vs the cross-encoder v0.5 and — for comparison only — Clef (clef-flash) and Jev (jev-latest):
| Family | kenning-xl v0.6 | kenning v0.5 | Clef | Jev |
|---|---|---|---|---|
| Macro | 0.688 | 0.653 | 0.791 | 0.830 |
| Agent (tool calls, task completion) | 0.860 | 0.727 | 0.793 | 0.900 |
| Conversation | 1.000 | 0.979 | 1.000 | 1.000 |
| Records (rules over JSON) | 0.663 | 0.587 | 0.857 | 0.921 |
| Tables | 0.660 | 0.520 | 0.860 | 0.940 |
| Text | 0.752 | 0.756 | 0.841 | 0.834 |
| Quality (helpful, correct) | 0.439 | 0.479 | 0.447 | 0.498 |
| Logs | 0.440 | 0.520 | 0.740 | 0.720 |
Beats Clef on agent decisions, ties conversation, and the gold-trace deliberate reasoning more than doubles the hardest arithmetic task (daily-limit checks 0.42 → 0.81, transferring to held-out tasks). It does not match Clef overall, and currently trails even the cross-encoder on logs, quality and text — deliberate reasoning isn't trace-trained for log-counting yet. Pick the engine for your task: Kenning-XL for agent/structured decisions, the v0.5 cross-encoder for text/safety/logs.
Licence
Weights under the Apache License 2.0 (merged LoRA adapter on Apache-2.0 Qwen3-1.7B). Training data is permissively licensed or generated; see NOTICE.md. Kenning implements a wire format compatible with TypeSafe AI's System One API; not affiliated with or endorsed by TypeSafe AI, and not trained on its outputs.