Remove summary paragraph

README.md

@@ -16,8 +16,6 @@ Models, attribution scores, and demo for **[Elucidating the Design Space of LM D
 
 <img src="https://huggingface.co/spaces/catt-bench/README/resolve/main/teaser.gif" alt="Token-level attribution under cross-entropy and negative log-odds" width="640">
 
-CATT fine-tunes Qwen3.5-4B on synthetic facts that are absent from pretraining, so the documents defining each fact are the only possible source of that knowledge and the attribution ground truth is known. Its datasets vary one property at a time: the number of documents per fact, the number of facts, and lexical distractors that mimic the query. Standard attribution methods dilute the few tokens that carry the evidence by averaging over the whole document. Scoring with negative log-odds gradients and keeping only the top-scoring tokens improves recall by 35 to 72 percentage points across methods.
-
 ## Models
 
 | Task | Base model (added tokens, not fine-tuned) | Fine-tuned LoRA adapters |