Line-level OCR for Amharic and Ge'ez: Surya fine-tuned on modern Amharic (Fidel), then on 18th–20th century manuscripts (HHD-Ethiopic).
Kidist Amde Mekonnen
kiyam
AI & ML interests
AI (Generative models, Computer vision, NLP) ,XAI
Recent Activity
updated a model 20 days ago
kiyam/surya-fidel-hhd-ethiopic upvoted a collection 20 days ago
DSpark & DFlash2 updated a collection 20 days ago
Ethiopic OCR: Amharic & Ge'ez Text RecognitionOrganizations
Step-3 DDRO optimized checkpoints (final policy) + accompanying datasets/artifacts (docIDs, pseudo-queries, testsets) to reproduce the paper.
Line-level OCR for Amharic and Ge'ez: Surya fine-tuned on modern Amharic (Fidel), then on 18th–20th century manuscripts (HHD-Ethiopic).
Preprocessed datasets for training and evaluating DDRO generative retrieval models on MS MARCO and Natural Questions.
Step-2 SFT reference policies (π_ref) used to initialize DDRO (MS MARCO / NQ; PQ and Title+URL DocIDs); use these for fair comparisons/ablations.
Step-3 DDRO optimized checkpoints (final policy) + accompanying datasets/artifacts (docIDs, pseudo-queries, testsets) to reproduce the paper.