MohammadKhosravi/llama3.1-8b-cpmt-cefr-balanced-topic-aligned

Balanced Topic-Aligned C-PMT Global Controller

This experiment evaluates the impact of incorporating explicit topic_title prompt metadata during C-PMT training on a balanced subset of EFCAMDAT (~5.5K total samples).

Experimental Setup

1. Topic Alignment: Prompt contexts incorporate topic_title directly, allowing the frozen base model to handle semantic content while the C-PMT controller isolates structural CEFR steering.
2. Blind Level Control: The text prompt contains no target CEFR text instructions. Control is enforced exclusively via the continuous cefr_id architectural knob.
3. Loss: Standard cross-entropy causal language modeling loss on balanced target tokens.