Update config: do_sample, max_new_tokens, top_p, temperature, repetition_penalty

config.json

@@ -119,7 +119,7 @@
       "type": "audio"
     }
   },
-  "do_sample": true,
+  "do_sample": false,
   "downsample_rate": 5,
   "dtype": "bfloat16",
   "enable_thinking": true,
@@ -252,7 +252,7 @@
     "up_proj",
     "down_proj"
   ],
-  "max_new_tokens": 128,
+  "max_new_tokens": 4096,
   "min_new_tokens": 0,
   "model_dtype": "bfloat16",
   "model_type": "asr_model",
@@ -275,10 +275,10 @@
   "qformer_num_heads": 16,
   "qformer_num_layers": 2,
   "qformer_window_size": 15,
-  "repetition_penalty": 1.1,
+  "repetition_penalty": 1.0,
   "router_aux_loss_coef": 0.01,
   "system_prompt": "",
-  "temperature": 1.0,
+  "temperature": 0.8,
   "text_config": {
     "_name_or_path": "HuggingFaceTB/SmolLM3-3B",
     "architectures": [
@@ -393,10 +393,10 @@
   "text_model_id": "HuggingFaceTB/SmolLM3-3B",
   "time_mask_length": 100,
   "top_k": 0,
-  "top_p": 0.01,
+  "top_p": 0.9,
   "transformers_version": "5.0.0",
   "use_cache": false,
   "use_lora": false,
   "use_specaugment": true,
   "vocab_size": 128257
-}
+}