Update README.md

README.md

@@ -178,7 +178,7 @@ From EmbeddingGemma-300m checkpoint; contrastive framework (InfoNCE with hard ne
 ## Evaluation
 
 ### Framework
-Assessed on internal multilingual retrieval suites (20-35% gains in low-resource hit@10). External benchmarks (MTEB multilingual, BEIR, Mr. TyDi, mMARCO, STS-B, XNLI, MultiGLUE) in preparation—projected 25% uplift.
+Assessed on internal multilingual retrieval suites (5-10% gains in low-resource hit@10). External benchmarks (MTEB multilingual, BEIR, Mr. TyDi, mMARCO, STS-B, XNLI, MultiGLUE) in preparation.
 
 Qualitative: Tight t-SNE clustering for parallels; excels in complex/mixed-language inputs.