feat: enhance LocalNMTService to set target language and log sample outputs in TranslationService

services/local_nmt_service.py

@@ -178,6 +178,8 @@ class LocalNMTService:
         """Runs a small mini-batch chunk on GPU using standard NLLB parameters."""
         if hasattr(self.tokenizer, "src_lang"):
             self.tokenizer.src_lang = src_code
+        if hasattr(self.tokenizer, "tgt_lang"):
+            self.tokenizer.tgt_lang = tgt_code
 
         # If on ZeroGPU and CUDA is now available, ensure model parameters are on CUDA
         if torch.cuda.is_available() and self.model:
@@ -193,17 +195,25 @@ class LocalNMTService:
         input_ids = inputs["input_ids"].to(target_device)
         attention_mask = inputs["attention_mask"].to(target_device)
 
+        # Retrieve exact target language token ID for decoder
+        target_bos = None
+        if hasattr(self.tokenizer, "convert_tokens_to_ids"):
+            try:
+                target_bos = self.tokenizer.convert_tokens_to_ids(tgt_code)
+            except Exception:
+                pass
+        if target_bos is None or target_bos < 0:
+            target_bos = forced_bos_id
+
         gen_kwargs = {
             "input_ids": input_ids,
             "attention_mask": attention_mask,
             "max_new_tokens": 256,
             "num_beams": 4,
-            "no_repeat_ngram_size": 3,
-            "repetition_penalty": 1.15,
             "early_stopping": True
         }
-        if forced_bos_id is not None:
-            gen_kwargs["forced_bos_token_id"] = forced_bos_id
+        if target_bos is not None:
+            gen_kwargs["forced_bos_token_id"] = target_bos
 
         with torch.no_grad():
             generated_tokens = self.model.generate(**gen_kwargs)
@@ -243,7 +253,8 @@ class LocalNMTService:
                 # Process this small chunk safely within ZeroGPU decorator limit
                 chunk_outputs = self._generate_chunk_on_gpu(batch_chunk, src_code, tgt_code, forced_bos_id)
                 chunk_dur = time.time() - chunk_start
-                logger.info(f"NLLB Batch chunk [{i}:{i+len(chunk)}] translated {len(chunk_outputs)} items in {chunk_dur:.2f}s")
+                sample_txt = chunk_outputs[0] if chunk_outputs else ""
+                logger.info(f"NLLB Batch chunk [{i}:{i+len(chunk)}] translated {len(chunk_outputs)} items in {chunk_dur:.2f}s. Sample: '{sample_txt[:70]}'")
                 results.extend(chunk_outputs)
             except Exception as chunk_err:
                 logger.warning(f"NLLB Batch chunk [{i}:{i+len(chunk)}] failed: {chunk_err}. Filling with None for controlled fallback.")

services/translation_service.py

@@ -392,13 +392,18 @@ class TranslationService:
                     t_nmt = time.time()
                     batch_outputs = local_nmt_service.translate_batch(batch_texts, target_lang=target_lang)
                     dur_nmt = time.time() - t_nmt
+                    if batch_outputs and len(batch_outputs) > 0 and batch_outputs[0]:
+                        logger.info(f"🔍 NLLB Sample Output [0]: '{batch_outputs[0][:80]}'")
                     for f_idx, out in zip(uncached_indices, batch_outputs):
-                        if out and contains_target_script(out, target_lang):
+                        is_valid = contains_target_script(out, target_lang) if out else False
+                        if is_valid:
                             cleaned_t = sanitize_translation_text(out)
                             raw_translations[f_idx] = cleaned_t
                             self._cache[f"{target_lang}:{merged_texts[f_idx].strip().lower()}"] = cleaned_t
                             provider_stats["local_nmt"] += 1
                             provider_timing["local_nmt"] += (dur_nmt / max(1, len(uncached_indices)))
+                        else:
+                            logger.warning(f"⚠️ NLLB output rejected for group [{f_idx}]: text='{out[:60] if out else '<empty/None>'}'")
                 except Exception as nmt_err:
                     logger.warning(f"Local NMT batch notice: {nmt_err}")