13 days ago
-
Model card: Ettin-150M is the base model (fine-tuned); SigLIP 2 is used frozen, credited in NOTICE.md
-
Model card: drop the distillation tag
-
Model card: tighter prose throughout; same facts
-
Model card: charts, MIT licence, a pass over the text; topic_classifier.py: parallel picture resize, one shared GPU worker, optional stage timing
Charts drawn by trainer/analysis/card_charts.py in the topic-feed project, from metrics.json and the test probabilities. -
Module decodes pictures on threads and prepares the next batch while the GPU works
-
Module reads picture bytes and reports pictures_used; config records the post document version
14 days ago
-
Fusion topic classifier: Ettin-150M + frozen SigLIP 2 picture embeddings, taxonomy v2.1
-
initial commit
