adjust the generation parameters to avoid repetitions

app.py

@@ -69,6 +69,9 @@ def _translate(text: str, src_lang: str, tgt_lang: str):
                 forced_bos_token_id=tokenizer.convert_tokens_to_ids(code_mapping[tgt_lang]),
                 max_length=len(input_tokens) + 50,
                 num_return_sequences=1,
+                num_beams=5,
+                no_repeat_ngram_size=4,  # repetition blocking works better if this number is below num_beams  
+                renormalize_logits=True,  # recompute token probabilities after banning the repetitions
             )
             translated_chunk = tokenizer.decode(
                 translated_chunk[0], skip_special_tokens=True