Back to browse
SRT-Adapter: 0.99 AUROC, perplexity win and 16.7× recall (0.19%)

SRT-Adapter: 0.99 AUROC, perplexity win and 16.7× recall (0.19%)

by spacebacon·Apr 26, 2026·2 points·0 comments

AI Analysis

MidNiche GemBold Bet

Training code withheld during patent review limits reproducibility and community adoption.

Strengths
  • 0.19% trainable parameters (14.5M) on frozen 7B backbone with no degradation
  • 0.99 AUROC on regime detection with 16.7× recall improvement over baseline
  • Four semiotic channel readouts per token position for discourse analysis
Weaknesses
  • Training pipelines and loss code not included during patent and publication review
  • Requires custom SRTAdapter class loader, not AutoModel-loadable checkpoint
Category
Target Audience

ML researchers, NLP engineers studying model interpretability

Similar To

LoRA · QLoRA · AdapterHub

Similar Projects

AI/ML●●●Banger

MicroGPT-C – C99 GPT for Edge Training and Tiny Model Pipelines

Karpathy's microgpt in C99, proves tiny coordinated models beat single large models on logic.

WizardryBig Brain
Ajay__soni
103mo ago