Merge pull request #17 from bencondemi/bc/dev

changing test suite

requirements.txt

@@ -1,10 +1,4 @@
 Flask==3.0.0
 pytest==7.4.0
 openai
-torch>=1.6.0
-transformers>=4.6.0
-numpy
-scipy
-scikit-learn
-tqdm
-nltk
+

src/translator.py

@@ -11,6 +11,7 @@
 
 
 
+
 def get_translation(content: str) -> str:
     """
     Translates non-English text to English using the Azure OpenAI GPT-4 model.

test/unit/eval_fns.py

@@ -42,44 +42,3 @@
 #     # Return average score
 #     return ((classification_score + translation_score) / 2)
 
-
-from transformers import AutoTokenizer, AutoModel
-import torch
-import numpy as np
-from scipy.spatial.distance import cosine
-
-# Load tokenizer and model from Hugging Face
-tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")
-model = AutoModel.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")
-
-def encode_text(text: str) -> np.ndarray:
-    # Tokenize input and convert to model embeddings
-    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
-    with torch.no_grad():
-        embeddings = model(**inputs).last_hidden_state.mean(dim=1).squeeze().numpy()
-    return embeddings
-
-def eval_single_response_translation(expected_answer: str, llm_response: str) -> float:
-    embeddings_1 = encode_text(expected_answer)
-    embeddings_2 = encode_text(llm_response)
-    # Cosine similarity between embeddings
-    similarity_score = 1 - cosine(embeddings_1, embeddings_2)
-    return similarity_score
-
-def eval_single_response_classification(expected_answer: str, llm_response: str) -> float:
-    return 1.0 if expected_answer.lower() == llm_response.strip().lower() else 0.0
-
-def eval_single_response_complete(expected_answer: tuple[bool, str], llm_response: tuple[bool, str]) -> float:
-    expected_is_english, expected_translation = expected_answer
-    response_is_english, response_translation = llm_response
-
-    # Evaluate classification
-    classification_score = eval_single_response_classification(expected_translation, response_translation)
-
-    # Evaluate translation only if not English
-    if not expected_is_english:
-        translation_score = eval_single_response_translation(expected_translation, response_translation)
-    else:
-        translation_score = eval_single_response_translation(expected_translation, response_translation)
-
-    return (classification_score + translation_score) / 2

test/unit/test_translator.py

@@ -1,35 +1,42 @@
 from src.translator import translate_content
-from test.unit.eval_sets import *
-from test.unit.eval_fns import eval_single_response_complete
+# from test.unit.eval_sets import *
+# from test.unit.eval_fns import eval_single_response_complete
 
 
 
-def test_chinese():
+# def test_chinese():
     # is_english, translated_content = translate_content("这是一条中文消息")
 
     # assert is_english == False
     # assert translated_content == "This is a Chinese message"
-    test = True
-    assert test == True
+
 
-eval_example_good = {"post": "Hier ist dein erstes Beispiel.", "expected_answer": (False, "Here is your first example.")}
-eval_example_bad = {"post": "asdfghjkl", "expected_answer": (False, "I don't understand your request.")}
+# eval_example_good = {"post": "Hier ist dein erstes Beispiel.", "expected_answer": (False, "Here is your first example.")}
+# eval_example_bad = {"post": "asdfghjkl", "expected_answer": (False, "I don't understand your request.")}
 
 
 def test_valid():
-    content = eval_example_good["post"]
-    expected = eval_example_good["expected_answer"]
-    llm_response = translate_content(content)
+    is_english, translated_content = translate_content("这是一条中文消息")
+
+    assert is_english == False
+    assert translated_content in ["This is a Chinese message.", 
+                                    "This is a message in Chinese."]
+    # content = eval_example_good["post"]
+    # expected = eval_example_good["expected_answer"]
+    # llm_response = translate_content(content)
 
-    similarity = eval_single_response_complete(expected, llm_response)
+    # similarity = eval_single_response_complete(expected, llm_response)
 
-    assert (0.90 <= similarity)
+    # assert (0.90 <= similarity)
 
 def test_invalid():
-    content = eval_example_bad["post"]
-    expected = eval_example_bad["expected_answer"]
-    llm_response = translate_content(content)
+    # content = eval_example_bad["post"]
+    # expected = eval_example_bad["expected_answer"]
+    # llm_response = translate_content(content)
 
+    is_english, translated_content = translate_content("asldjkjhfjsdh")
+
+    assert is_english == False
     assert ValueError("Invalid translation response.")