martini/laya-query-router

Laya (Fine-Tuned Query Router)

Fine-tuned on a local dataset of 500 template-generated queries with multi-hot labels over 32 crops, 6 roles, and 19 zones. The checkpoint drops into the existing Laya Router/Agent two-step call pattern: step 1 asks one noul presence question per category, step 2 asks one noul question per option for gated categories, with membership decisions taken at the per-option thresholds stored in rl_agent_config.json (thresholds) rather than a global 0.5.

Evaluation (held-out 100-query test split)

Model	Step-1 Acc (mean)	Step-2 P/R/F1	Label-Set F1	Latency p50 (ms)
Laya (fine-tuned, gated)	1.0	0.959 / 1.000 / 0.979	0.9893	173.0
Laya (fine-tuned, always-on)	1.0	0.946 / 1.000 / 0.972	0.9693	252.8
Laya (pretrained, zero-shot)	0.713	nan / nan / nan	nan	46.7

Limitations

• Template leakage: queries share ~100 sentence frames; absolute metrics partly measure template memorization.
• Multi-label supervision is scarce (10 multi-crop, 5 multi-zone queries), so multi-label metrics are indicative rather than decisive.

Installation & Quickstart

pip install laya

import laya

agent = laya.load("convaiinnovations/laya-query-router")
res = agent.predict(state, questions)
print(res["answers"])

License

Apache 2.0. Developed by Convai Innovations.