🪔 Doha Generation Model v3

What's New in v3 (Stage 2 v5)

• Gate regularization — balanced encoder/meaning attention
• Meaning decoder pre-training phase — 10 epochs before joint training
• Enriched meaning with theme prefix विषय: {theme} | {meaning}
• Encoder unfreeze bug fixed — relative to START_EPOCH
• Early stopping — prevents overfitting past peak performance
• Matra best-of-N generation — picks best syllable structure

Architecture

• Shared Encoder pretrained on 58k Kavitas via T5-style span corruption
• Meaning Decoder generates semantic meaning first
• Doha Decoder uses dual cross attention (encoder + meaning) with learnable gate

Usage