drop initial_prompt (rejected by transformers 4.46 generate kwargs)

kaggle/ming_eval.py

@@ -174,12 +174,13 @@ def build_asr():
         pipe = pipeline("automatic-speech-recognition", torch_dtype=torch.float16, **kwargs)
     except Exception:
         pipe = pipeline("automatic-speech-recognition", **kwargs)
-    prompt = "以下是普通话的句子，请使用简体中文转写。"  # pushes Whisper off traditional output
-
     def asr(wav: np.ndarray, sr: int = 44100) -> str:
+        # No initial_prompt here: transformers 4.46 rejects it as a generate
+        # kwarg. Whisper's traditional-character output is handled downstream
+        # by eval/score.py, which normalises before counting edits.
         out = pipe(
             {"array": wav.astype(np.float32), "sampling_rate": sr},
-            generate_kwargs={"language": "zh", "task": "transcribe", "initial_prompt": prompt},
+            generate_kwargs={"language": "zh", "task": "transcribe"},
         )
         return out.get("text", "") if isinstance(out, dict) else str(out)