SGK86/laya-multilingual-typed-decisions

Laya-multilingual fine-tuned on LocalLLaMA/typed-decisions (RLCD)

Fine-tuned from convaiinnovations/laya-multilingual with the official RLCD recipe (single GPU, effective batch 64, lr encoder 2.5e-5 / head 1e-4, 4 epochs, sigma 0.4->0.1), with post-training temperature calibration.

Official 400-case test set metrics: Accuracy 0.786, Soft Acc 0.507, Brier 0.064, ECE 0.113, Score MAE 0.253.

Reference: TypeSafe Jev 1.13.0 (third-party, BENCHMARKS.md): Acc 0.727, Soft Acc 0.580, Brier 0.148, ECE 0.144, MAE 0.391. Majority baseline: 0.461.

import laya
agent = laya.load("SGK86/laya-multilingual-typed-decisions")
result = agent.predict(state, questions)