Merge branch 'UKPLab:master' into Prompting-on-evaluators

UKPLab · Sep 23, 2024 · f95cb46 · f95cb46
2 parents 005039f + 7290448
commit f95cb46
Show file tree

Hide file tree

Showing 2 changed files with 18 additions and 7 deletions.
diff --git a/pyproject.toml b/pyproject.toml
@@ -1,12 +1,15 @@
 [project]
 name = "sentence-transformers"
 version = "3.2.0.dev0"
-description = "Multilingual text embeddings"
-license = { file = "LICENSE" }
+description = "State-of-the-Art Text Embeddings"
+license = { text = "Apache 2.0" }
 readme = "README.md"
 authors = [
     { name = "Nils Reimers", email = "info@nils-reimers.de" },
-    { name = "Tom Aarsen" },
+    { name = "Tom Aarsen", email = "tom.aarsen@huggingface.co" },
+]
+maintainers = [
+    { name = "Tom Aarsen", email = "tom.aarsen@huggingface.co" }
 ]
 requires-python = ">=3.8"
 keywords = [

diff --git a/sentence_transformers/util.py b/sentence_transformers/util.py
@@ -714,8 +714,12 @@ def mine_hard_negatives(
         except Exception:
             pass
 
-        corpus_embeddings = model.encode(corpus, batch_size=batch_size, convert_to_numpy=True, show_progress_bar=True)
-        query_embeddings = model.encode(queries, batch_size=batch_size, convert_to_numpy=True, show_progress_bar=True)
+        corpus_embeddings = model.encode(
+            corpus, batch_size=batch_size, normalize_embeddings=True, convert_to_numpy=True, show_progress_bar=True
+        )
+        query_embeddings = model.encode(
+            queries, batch_size=batch_size, normalize_embeddings=True, convert_to_numpy=True, show_progress_bar=True
+        )
         index.add(corpus_embeddings)
 
         scores_list = []
@@ -731,8 +735,12 @@ def mine_hard_negatives(
 
     else:
         # Embed the corpus and the queries
-        corpus_embeddings = model.encode(corpus, batch_size=batch_size, convert_to_numpy=True, show_progress_bar=True)
-        query_embeddings = model.encode(queries, batch_size=batch_size, convert_to_numpy=True, show_progress_bar=True)
+        corpus_embeddings = model.encode(
+            corpus, batch_size=batch_size, normalize_embeddings=True, convert_to_numpy=True, show_progress_bar=True
+        )
+        query_embeddings = model.encode(
+            queries, batch_size=batch_size, normalize_embeddings=True, convert_to_numpy=True, show_progress_bar=True
+        )
         scores = model.similarity(query_embeddings, corpus_embeddings).to(device)
 
         # Keep only the range_max + max_positives highest scores. We offset by 1 to potentially include the positive pair