This video shows my Tetris test of Contrastive Language Models (CLM). CLM uses contrastive learning to match a state with candidate actions. The released CLM-8B model combines Qwen3-8B embeddings with learned projection heads to score these actions.

Download the video (.mov).

CLM reached 82.90% accuracy in my IMDb movie review classification test and failed the Tetris test shown here. I discuss these results and the broader classification context in Language Models for Text Classification: From Bag-of-Words to Jev.