Laya MoE expert heads

Domain expert heads for layaMOE. Each folder holds one decision head (type embedding, 2 transformer layers, scorer; 26.5M parameters) fine-tuned from the head of convaiinnovations/laya-typed-decisions on synthetic, rule-labeled domain data, with its own calibration temperatures in expert.json. The encoder is not included: load it from the base checkpoint, which is shared by all experts.

Expert	Domains	Validation acc. (base head -> expert)
`customer_ops`	delivery-exceptions, email-triage, support-tickets	57.3% -> 75.9%
`safety`	agent-guardrails, content-moderation	58.4% -> 73.4%

from laya_moe.moe import MoEAgent   # github.com/vishalmysore/layaMOE
from huggingface_hub import snapshot_download
d = snapshot_download("VishalMysore/layaMOE")
moe = MoEAgent(experts=[f"{d}/safety", f"{d}/customer_ops"])

Modified derivative of laya-typed-decisions (Apache-2.0, Copyright ConvAI Innovations); unofficial and not affiliated with ConvAI Innovations. The training data is synthetic; evaluate on your own cases before relying on it. See LICENSE and NOTICE.md.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for VishalMysore/layaMOE

Base model

convaiinnovations/laya-typed-decisions

Finetuned

(4)

this model