Skip to content
Igor Ferreira

Evidence panel

laya-tema, the topic router for the Mesa

CPU fine-tuning of the multilingual Laya model (Apache 2.0) to choose which of the 9 topics of Mesa dos 28 a question belongs to, with measured calibration and abstention below a confidence threshold.

Track
Machine learning and data
Role
Author, personal project
Period
2026
Status
in progress
Tools
Python, Laya multilingual, safetensors

Evidence

  • 85% accuracy on the locked final set (ECE 0.08)

    108 questions, evaluated a single time on October 11, 2026. On the same test, the keyword rule got 50% right.

    Measured.

  • 96.9% accuracy covering 59% of the questions

    Confidence threshold τ = 0.8. τ was set after seeing the final set, so the next version needs a new final set.

    Measured.

Limitations

The sets were written and labeled by different subagents, and the labels have not yet been reviewed by a person. Questions without a keyword are the weak spot and the next training data.