Training in progress - step 23000

asr_config.py

@@ -25,6 +25,7 @@ class ASRConfig(transformers.PretrainedConfig):
         model_dtype: str = "bfloat16",
         num_beams: Optional[int] = None,
         system_prompt: str = "You are a helpful assistant.",
+        user_prompt: str = "Please transcribe this English audio into text: <audio>",
         encoder_dim: Optional[int] = None,
         llm_dim: Optional[int] = None,
         # Encoder conv layers: list of (padding, kernel_size, stride) tuples
@@ -103,6 +104,7 @@ class ASRConfig(transformers.PretrainedConfig):
         self.attn_implementation = attn_implementation
         self.model_dtype = model_dtype
         self.system_prompt = system_prompt
+        self.user_prompt = user_prompt
         self.encoder_dim = encoder_dim
         self.llm_dim = llm_dim
         # Default conv layers for Whisper/GLM-ASR: [(pad, kernel, stride), ...]
@@ -204,10 +206,6 @@ class ASRConfig(transformers.PretrainedConfig):
 
         super().__init__(**kwargs)
 
-        # Point encoder to audio_config so pipeline uses correct feature extractor
-        # The pipeline looks for config.encoder._name_or_path for feature extractor
-        self.encoder = self.audio_config
-
         self.auto_map = {
             "AutoConfig": "asr_config.ASRConfig",
             "AutoModel": "asr_modeling.ASRModel",