Create app.py

app.py

@@ -0,0 +1,84 @@
+import streamlit as st
+from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
+import evaluate
+
+# Page configuration
+st.set_page_config(
+    page_title="Translation Evaluator",
+    layout="wide",
+    initial_sidebar_state="collapsed"
+)
+
+# Load model and tokenizer
+@st.cache_resource
+def load_model():
+    model_name = "facebook/nllb-200-distilled-600M"
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
+    return tokenizer, model
+
+tokenizer, model = load_model()
+
+# Load metrics
+bleu = evaluate.load("bleu")
+bertscore = evaluate.load("bertscore")
+comet = evaluate.load("comet", module_type="metric")
+# For BERTurk, use Turkish BERT for BERTScore
+bertturk = evaluate.load("bertscore")
+
+# UI
+st.title("Minimalistic Translation & Evaluation")
+input_text = st.text_area("Input text (any language)", height=150)
+ref_text = st.text_area("Reference translation in Turkish (optional)", height=150)
+
+if st.button("Translate & Evaluate"):
+    if not input_text.strip():
+        st.error("Please enter some input text to translate.")
+    else:
+        # Tokenize and generate
+        inputs = tokenizer(input_text, return_tensors="pt")
+        outputs = model.generate(**inputs, forced_bos_token_id=tokenizer.get_lang_id("tur_TUR"))
+        translation = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
+
+        # Display translation
+        st.subheader("Model Translation (Turkish):")
+        st.write(translation)
+
+        # Prepare references and predictions
+        predictions = [translation]
+        references = [[ref_text]] if ref_text.strip() else None
+
+        # Compute BLEU
+        if references:
+            bleu_result = bleu.compute(predictions=predictions, references=references)
+            st.metric("BLEU-4", f"{bleu_result['bleu']*100:.2f}")
+        else:
+            st.info("No reference provided: skipping BLEU.")
+
+        # Compute BERTScore (multilingual)
+        bs = bertscore.compute(
+            predictions=predictions,
+            references=[ref_text] if ref_text.strip() else [translation],
+            lang="tr"
+        )
+        st.metric("BERTScore (f1)", f"{bs['f1'][0]*100:.2f}")
+
+        # Compute BERTurk specifically
+        bt = bertturk.compute(
+            predictions=predictions,
+            references=[ref_text] if ref_text.strip() else [translation],
+            model_type="dbmdz/bert-base-turkish-cased"
+        )
+        st.metric("BERTurk (f1)", f"{bt['f1'][0]*100:.2f}")
+
+        # Compute COMET if reference
+        if references:
+            comet_score = comet.compute(
+                model="Unbabel/wmt22-comet-da",
+                src=[input_text],
+                mt=predictions,
+                ref=[ref_text]
+            )
+            st.metric("COMET score", f"{comet_score['score'][0]:.2f}")
+        else:
+            st.info("No reference provided: skipping COMET.")