MohammadKhosravi/llama3.1-8b-cpmt-cefr-wce-topic-aligned-60k

60K Topic-Aligned C-PMT Global Controller (Weighted Cross Entropy Edition)

This controller is trained across the full 60K EFCAMDAT corpus using Topic-Aligned Prompts and Weighted Cross-Entropy Loss.

Experimental Setup

1. Topic Alignment: Incorporates topic_title directly into the prompt context, allowing the base LLM to handle semantics while the C-PMT controller isolates structural CEFR steering.
2. Blind Level Control: The prompt text contains no target CEFR text string. Control is enforced exclusively via the continuous cefr_id architectural knob.
3. Loss Function: Sequence-level Weighted Cross-Entropy (WCE) loss to balance gradient flow across minority classes (C1/C2).