Update README.md

README.md

@@ -37,7 +37,7 @@ The training followed a multilingual knowledge distillation approach introduced
 
 To evaluate clustering quality a Polish clustering benchmark was created with eight datasets covering short texts, long documents, web pages, news articles, scientific abstracts and domain-specific banking content. The benchmark combines existing Polish tasks with datasets translated from English to Polish using an LLM. The included tasks are: **8TAGS** (clustering headlines into 8 topic categories), **Banking-Long** (clustering longer finance and banking related articles into 14 topical sections), **Banking-Short** (clustering finance and banking headlines into the same 14 topics), **CORTEX** (clustering short texts by expressed emotion into 9 classes), **Fineweb-Topics** (clustering Polish web pages into 22 topic categories), **Guardian-Articles** (clustering translated Guardian articles into 29 editorial sections), **PLSC** (clustering scientific titles and abstracts by research field into 8 classes) and **WikiNewsPL** (clustering WikiNews articles into 9 news topics).
 
-The evaluation results for the new models together with comparisons to popular Polish and multilingual embedding models are presented in Tables 1 and 2. For each model, documents were encoded into vector representations and clustered with K-Means algorithm. Table 1 reports clustering quality using ARI (Adjusted Rand Index), while Table 2 uses V-measure. As a reference point, the evaluation also includes a  lexical baseline method based on TF-IDF representations reduced to dense vectors with Truncated SVD. This baseline represents texts using word-level lexical information rather than learned semantic embeddings.
+The evaluation results for the new models together with comparisons to popular Polish and multilingual embedding models are presented in Tables 1 and 2. For each model, documents were encoded into vector representations and clustered with K-Means algorithm. Table 1 reports clustering quality using ARI (Adjusted Rand Index), while Table 2 uses V-measure. Each clustering experiment was repeated five times and the results reported in below tables are average ARI/V-measure scores. As a reference point, the evaluation also includes a  lexical baseline method based on TF-IDF representations reduced to dense vectors with Truncated SVD. This baseline represents texts using word-level lexical information rather than learned semantic embeddings.
 
 <table>
     <thead>