Property	Value
Base model	`NeuML/pubmedbert-base-embeddings`
Output dimension	256
Training method	Triplet-loss contrastive learning
Training corpus	563,845 ClinicalTrials.gov entries
Grouping signal	10,609 MeSH intervention terms
Epochs	10
Batch size	32
Learning rate	`2e-5`
Loss margin	0.5
Training hardware	NVIDIA DGX Spark, GB10 Blackwell

Intended use

MOAt supports similarity analysis by biological mechanism. It does not estimate treatment efficacy or clinical-trial success.

Architecture

Trial text -> PubMedBERT embedding -> dense 384-to-256 projection -> L2-normalized vector

Limitations

The training signal depends on ClinicalTrials.gov records and MeSH intervention assignments. Missing or inconsistent records can affect similarity results.

This repository contains the original published model card and configuration scope. The source repository did not publish executable weight files.

AcademicaMed family

PACT: population and condition similarity.
FATE: historical outcome-pattern similarity.
ORACLE: calibrated synthesis classifier.

Academica/AcademicaMed-MOAt-ClinicalTrials-MoA-256

AcademicaMed-MOAt-ClinicalTrials-MoA-256

Model details

Intended use

Architecture

Limitations

AcademicaMed family

Academica/
AcademicaMed-MOAt-ClinicalTrials-MoA-256