fix: fixing model config settings

explanation/attention.py

@@ -15,7 +15,7 @@ def chat_explained(model, prompt):
     ).input_ids
     # generate output together with attentions of the model
     decoder_input_ids = model.MODEL.generate(
-        encoder_input_ids, output_attentions=True, **model.CONFIG
+        encoder_input_ids, output_attentions=True, generation_config=model.CONFIG
     )
 
     # get input and output text as list of strings

main.py

@@ -110,9 +110,10 @@ with gr.Blocks(
                         label="System Prompt",
                         info="Set the models system prompt, dictating how it answers.",
                         # default system prompt is set to this in the backend
-                        placeholder=(
-                            "You are a helpful, respectful and honest assistant. Always"
-                            " answer as helpfully as possible, while being safe."
+                        placeholder=("""
+                            You are a helpful, respectful and honest assistant. Always
+                            answer as helpfully as possible, while being safe.
+                            """
                         ),
                     )
                 # column that takes up 1/4 of the row
@@ -121,7 +122,7 @@ with gr.Blocks(
                     xai_selection = gr.Radio(
                         ["None", "SHAP", "Attention"],
                         label="Interpretability Settings",
-                        info="Select a Interpretability Implementation to use.",
+                        info="Select a Interpretability Approach Implementation to use.",
                         value="None",
                         interactive=True,
                         show_label=True,
@@ -133,15 +134,15 @@ with gr.Blocks(
                         ["GODEL", "Mistral"],
                         label="Model Settings",
                         info="Select a Model to use.",
-                        value="GODEL",
+                        value="Mistral",
                         interactive=True,
                         show_label=True,
                     )
 
                 # calling info functions on inputs/submits for different settings
-                system_prompt.submit(system_prompt_info, [system_prompt])
-                xai_selection.input(xai_info, [xai_selection])
-                model_selection.input(model_info, [model_selection])
+                system_prompt.change(system_prompt_info, [system_prompt])
+                xai_selection.change(xai_info, [xai_selection])
+                model_selection.change(model_info, [model_selection])
 
         # row with chatbot ui displaying "conversation" with the model
         with gr.Row(equal_height=True):

model/godel.py

@@ -1,7 +1,7 @@
 # GODEL model module for chat interaction and model instance control
 
 # external imports
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, GenerationConfig
 
 # internal imports
 from utils import modelling as mdl
@@ -10,24 +10,20 @@ from utils import modelling as mdl
 TOKENIZER = AutoTokenizer.from_pretrained("microsoft/GODEL-v1_1-large-seq2seq")
 MODEL = AutoModelForSeq2SeqLM.from_pretrained("microsoft/GODEL-v1_1-large-seq2seq")
 
-# default model config
-CONFIG = {"max_new_tokens": 50, "min_length": 8, "top_p": 0.9, "do_sample": True}
+
+# model config definition
+CONFIG = GenerationConfig.from_pretrained("microsoft/GODEL-v1_1-large-seq2seq")
+base_config_dict = {"max_new_tokens": 50, "min_length": 8, "top_p": 0.9, "do_sample": True}
+CONFIG.update(**base_config_dict)
 
 
 # function to (re) set config
-def set_config(config: dict):
-    global CONFIG
+def set_config(config_dict: dict):
 
-    # if config dict is given, update it
-    if config != {}:
-        CONFIG = config
-    else:
-        # hard setting model config to default
-        # needed for shap
-        MODEL.config.max_new_tokens = 50
-        MODEL.config.min_length = 8
-        MODEL.config.top_p = 0.9
-        MODEL.config.do_sample = True
+    # if config dict is not given, set to default
+    if config_dict == {}:
+        config_dict = base_config_dict
+    CONFIG.update(**config_dict)
 
 
 # formatting class to formatting input for the model
@@ -67,7 +63,7 @@ def respond(prompt):
     input_ids = TOKENIZER(f"{prompt}", return_tensors="pt").input_ids
 
     # generating using config and decoding output
-    outputs = MODEL.generate(input_ids, **CONFIG)
+    outputs = MODEL.generate(input_ids,generation_config=CONFIG)
     output = TOKENIZER.decode(outputs[0], skip_special_tokens=True)
 
     # returns the model output string