feat/fixing: correcting bug, updating documentation (final?)

README.md

@@ -21,7 +21,7 @@ This is the UI showcase for my thesis about the interpretability of LLM based ch
 
 ### 🔗 Links:
 
-**[Github Repository](https://github.com/LennardZuendorf/thesis-webapp)**
+**[GitHub Repository](https://github.com/LennardZuendorf/thesis-webapp)**
 
 **[Huggingface Spaces Showcase](https://huggingface.co/spaces/lennardzuendorf/thesis-webapp-docker)**
 
@@ -86,7 +86,7 @@ See code for in detailed credits, work is strongly based on:
 
 #### SHAP
 - [Github](https://github.com/shap/shap)
-- [Inital Paper](https://arxiv.org/abs/1705.07874)
+- [Initial Paper](https://arxiv.org/abs/1705.07874)
 
 #### Custom Component (/components/iframe/)
 

backend/controller.py

@@ -43,6 +43,7 @@ def explained_chat(
     #    message, history, system_prompt, knowledge
     # )
     prompt = model.format_prompt(message, history, system_prompt, knowledge)
+    print(f"Formatted prompt: {prompt}")
 
     # generating an answer using the methods chat function
     answer, xai_graphic, xai_markup, xai_plot = xai.chat_explained(model, prompt)
@@ -73,10 +74,10 @@ def interference(
     # if a model is selected, grab the model instance
     if model_selection.lower() == "mistral":
         model = mistral
-        print("Indentified model as Mistral")
+        print("Identified model as Mistral")
     else:
         model = godel
-        print("Indentified model as GODEL")
+        print("Identified model as GODEL")
 
     # if a XAI approach is selected, grab the XAI module instance
     # and call the explained chat function

components/iframe/README.md

@@ -1,7 +1,7 @@
 # gradio iFrame
 
 This is a custom gradio component used to display the shap package text plot. Which is interactive HTML and needs a custom wrapper.
-See custom component examples at offical [docu](https://www.gradio.app/guides/custom-components-in-five-minutes)
+See custom component examples at official [docu](https://www.gradio.app/guides/custom-components-in-five-minutes)
 
 # Credit
 CREDIT: based mostly of Gradio template component, HTML
@@ -14,4 +14,4 @@ see: https://www.gradio.app/docs/html
 - backend/iframe.py - updating component to accept custom height/width and added new example
 - demo/app.py - slightly changed demo file for better dev experience
 - frontend/index.svelte - slightly changed to accept custom height/width
-- frontend/HTML.svelte - updated to use iFrame and added custom function to programmtically set heigth values
+- frontend/HTML.svelte - updated to use iFrame and added custom function to programmatically set height values

explanation/attention.py

@@ -11,6 +11,8 @@ from .markup import markup_text
 # and marked text based on attention
 def chat_explained(model, prompt):
 
+    print(f"Running explained chat with prompt {prompt}.")
+
     # get encoded input
     input_ids = model.TOKENIZER(
         prompt, return_tensors="pt", add_special_tokens=True
@@ -56,6 +58,7 @@ def chat_explained(model, prompt):
         " Visualization doesn't support an interactive graphic.</h4></div>"
     )
     # creating marked text using markup_text function and attention
+    print(f"Creating marked text with {input_text}.")
     marked_text = markup_text(input_text, averaged_attention, variant="visualizer")
 
     # returning response, graphic and marked text array

explanation/interpret_captum.py

@@ -47,7 +47,7 @@ def chat_explained(model, prompt):
     # getting response text, graphic placeholder and marked text object
     response_text = fmt.format_output_text(attribution_result.output_tokens)
     graphic = """<div style='text-align: center; font-family:arial;'><h4>
-        Intepretation with Captum doesn't support an interactive graphic.</h4></div>
+        Interpretation with Captum doesn't support an interactive graphic.</h4></div>
         """
     # create the explanation marked text array
     marked_text = markup_text(input_tokens, values, variant="captum")

explanation/markup.py

@@ -21,7 +21,7 @@ def markup_text(input_text: list, text_values: ndarray, variant: str):
     elif variant == "visualizer":
         text_values = fmt.flatten_attention(text_values)
 
-    # Determine the minimum and maximum values
+    # determine the minimum and maximum values
     min_val, max_val = np.min(text_values), np.max(text_values)
 
     # separate the threshold calculation for negative and positive values
@@ -69,7 +69,7 @@ def color_codes():
     return {
         # -5 to -1: Strong Light Sky Blue to Lighter Sky Blue
         # 0: white (assuming default light mode)
-        # +1 to +5 light pink to strng magenta
+        # +1 to +5 light pink to strong magenta
         "-5": "#008bfb",
         "-4": "#68a1fd",
         "-3": "#96b7fe",

explanation/plotting.py

@@ -7,24 +7,24 @@ import matplotlib.pyplot as plt
 
 def plot_seq(seq_values: list, method: str = ""):
 
-    # Separate the tokens and their corresponding importance values
+    # separate the tokens and their corresponding importance values
     tokens, importance = zip(*seq_values)
 
-    # Convert importance values to numpy array for conditional coloring
+    # convert importance values to numpy array for conditional coloring
     importance = np.array(importance)
 
-    # Determine the colors based on the sign of the importance values
+    # determine the colors based on the sign of the importance values
     colors = ["#ff0051" if val > 0 else "#008bfb" for val in importance]
 
-    # Create a bar plot
+    # create a bar plot
     plt.figure(figsize=(len(tokens) * 0.9, np.max(importance)))
     x_positions = range(len(tokens))  # Positions for the bars
 
-    # Creating vertical bar plot
+    # creating vertical bar plot
     bar_width = 0.8
     plt.bar(x_positions, importance, color=colors, align="center", width=bar_width)
 
-    # Annotating each bar with its value
+    # annotating each bar with its value
     padding = 0.1  # Padding for text annotation
     for x, (y, color) in enumerate(zip(importance, colors)):
         sign = "+" if y > 0 else ""

main.py

@@ -26,7 +26,7 @@ css = """
     .examples {text-align: start;}
     .seperatedRow {border-top: 1rem solid;}",
     """
-# custom js to force lightmode in custom environments
+# custom js to force light mode in custom environments
 if os.environ["HOSTING"].lower() != "spaces":
     js = """
     function () {
@@ -52,6 +52,12 @@ def load_md(path):
 
 # function to display the system prompt info
 def system_prompt_info(sys_prompt_txt):
+    if sys_prompt_txt == "":
+        sys_prompt_txt = """
+            You are a helpful, respectful and honest assistant.
+            Always answer as helpfully as possible, while being safe.
+        """
+
     # display the system prompt using the Gradio Info component
     gr.Info(f"The system prompt was set to:\n {sys_prompt_txt}")
 
@@ -71,7 +77,7 @@ def model_info(model_radio):
 
 
 # ui interface based on Gradio Blocks
-# see https://www.gradio.app/docs/interface)
+# see https://www.gradio.app/docs/interface
 with gr.Blocks(
     css=css,
     js=js,
@@ -171,11 +177,11 @@ with gr.Blocks(
                     show_copy_button=True,
                     avatar_images=("./public/human.jpg", "./public/bot.jpg"),
                 )
-                # extenable components for extra knowledge
+                # extendable components for extra knowledge
                 with gr.Accordion(label="Additional Knowledge", open=False):
                     gr.Markdown("""
                         *Hint:* Add extra knowledge to see GODEL work the best.
-                        Knowledge doesn't work mith Mistral and will be ignored.
+                        Knowledge doesn't work with Mistral and will be ignored.
                         """)
                     # textbox to enter the knowledge
                     knowledge_input = gr.Textbox(
@@ -217,8 +223,8 @@ with gr.Blocks(
                             "Does money buy happiness?",
                             "",
                             (
-                                "Respond from the perspective of a billionaire enjoying"
-                                " life in Dubai"
+                                "Respond from the perspective of billionaire heir"
+                                " living his best life with his father's money."
                             ),
                             "Mistral",
                             "None",
@@ -227,8 +233,8 @@ with gr.Blocks(
                             "Does money buy happiness?",
                             "",
                             (
-                                "Respond from the perspective of a billionaire enjoying"
-                                " life in Dubai"
+                                "Respond from the perspective of billionaire heir"
+                                " living his best life with his father's money."
                             ),
                             "Mistral",
                             "SHAP",
@@ -251,14 +257,36 @@ with gr.Blocks(
                         [
                             "Does money buy happiness?",
                             (
-                                "Black holes are created when a massive star's core"
-                                " collapses after a supernova, forming an object with"
-                                " gravity so intense that even light cannot escape."
+                                "Some studies have found a correlation between income"
+                                " and happiness, but this relationship often has"
+                                " diminishing returns. From a psychological standpoint,"
+                                " it's not just having money, but how it is used that"
+                                " influences happiness."
                             ),
                             "",
                             "GODEL",
                             "SHAP",
                         ],
+                        [
+                            "Does money buy happiness?",
+                            (
+                                "Some studies have found a correlation between income"
+                                " and happiness, but this relationship often has"
+                                " diminishing returns. From a psychological standpoint,"
+                                " it's not just having money, but how it is used that"
+                                " influences happiness."
+                            ),
+                            "",
+                            "GODEL",
+                            "Attention",
+                        ],
+                        [
+                            "Does money buy happiness?",
+                            "",
+                            "",
+                            "GODEL",
+                            "Attention",
+                        ],
                     ],
                     inputs=[
                         user_prompt,
@@ -332,7 +360,7 @@ with gr.Blocks(
         # load about.md markdown
         gr.Markdown(value=load_md("public/about.md"))
         with gr.Accordion(label="Credits, Data Protection, License"):
-            # load credits and dataprotection markdown
+            # load credits and data protection markdown
             gr.Markdown(value=load_md("public/credits_dataprotection_license.md"))
 
 # mount function for fastAPI Application

model/godel.py

@@ -6,7 +6,7 @@ from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, GenerationConfig
 # internal imports
 from utils import modelling as mdl
 
-# global model and tokenizer instance (created on inital build)
+# global model and tokenizer instance (created on initial build)
 TOKENIZER = AutoTokenizer.from_pretrained("microsoft/GODEL-v1_1-large-seq2seq")
 MODEL = AutoModelForSeq2SeqLM.from_pretrained("microsoft/GODEL-v1_1-large-seq2seq")
 

model/mistral.py

@@ -9,7 +9,8 @@ import gradio as gr
 from utils import modelling as mdl
 from utils import formatting as fmt
 
-# global model and tokenizer instance (created on inital build)
+# global model and tokenizer instance (created on initial build)
+# determine if GPU is available and load model accordingly
 device = mdl.get_device()
 if device == torch.device("cuda"):
     n_gpus, max_memory, bnb_config = mdl.gpu_loading_config()
@@ -17,13 +18,15 @@ if device == torch.device("cuda"):
     MODEL = AutoModelForCausalLM.from_pretrained(
         "mistralai/Mistral-7B-Instruct-v0.2",
         quantization_config=bnb_config,
-        device_map="auto",  # dispatch efficiently the model on the available ressources
+        device_map="auto",
         max_memory={i: max_memory for i in range(n_gpus)},
     )
 
+# otherwise, load model on CPU
 else:
     MODEL = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
     MODEL.to(device)
+# load tokenizer
 TOKENIZER = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
 
 # default model config
@@ -48,12 +51,13 @@ def set_config(config_dict: dict):
     CONFIG.update(**config_dict)
 
 
-# advanced formatting function that takes into a account a conversation history
-# CREDIT: adapated from the Mistral AI Instruct chat template
+# advanced formatting function that takes into account a conversation history
+# CREDIT: adapted from the Mistral AI Instruct chat template
 # see https://github.com/chujiezheng/chat_templates/
 def format_prompt(message: str, history: list, system_prompt: str, knowledge: str = ""):
     prompt = ""
 
+    # send information to the ui if knowledge is not empty
     if knowledge != "":
         gr.Info("""
             Mistral doesn't support additional knowledge, it's gonna be ignored.
@@ -94,7 +98,7 @@ def format_answer(answer: str):
 
     # checking if proper history got returned
     if len(segments) > 1:
-        # return text after the last ['/INST'] - reponse to last message
+        # return text after the last ['/INST'] - response to last message
         formatted_answer = segments[-1].strip()
     else:
         # return warning and full answer if not enough [/INST] tokens found
@@ -108,7 +112,11 @@ def format_answer(answer: str):
     return formatted_answer
 
 
+# response class calling the model and returning the model output message
+# CREDIT: Copied from official interference example on Huggingface
+# see https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2
 def respond(prompt: str):
+    # setting config to default
     set_config({})
 
     # tokenizing inputs and configuring model
@@ -117,6 +125,9 @@ def respond(prompt: str):
     # generating text with tokenized input, returning output
     output_ids = MODEL.generate(input_ids, generation_config=CONFIG)
     output_text = TOKENIZER.batch_decode(output_ids)
+
+    # formatting output text with special function
     output_text = fmt.format_output_text(output_text)
 
+    # returning the model output string
     return format_answer(output_text)

utils/formatting.py

@@ -100,7 +100,7 @@ def avg_attention(attention_values, model: str):
 
     # removing the last dimension and transposing to get the correct shape
     attention = attention[:, :, :, 0]
-    attention = attention.transpose
+    attention = attention.transpose()
 
     # return the averaged attention values
     return np.mean(attention, axis=1)

utils/modelling.py

@@ -45,7 +45,7 @@ def prompt_limiter(
 
     # if token count small enough, adding history bit by bit
     if pre_count < 800:
-        # setting the count to the precount
+        # setting the count to the pre-count
         count = pre_count
         # reversing the history to prioritize recent conversations
         history.reverse()
@@ -76,6 +76,7 @@ def token_counter(tokenizer, text: str):
     return len(tokens[0])
 
 
+# function to determine the device to use
 def get_device():
     if torch.cuda.is_available():
         device = torch.device("cuda")
@@ -85,7 +86,9 @@ def get_device():
     return device
 
 
-# setting device based on available hardware
+# function to set device config
+# CREDIT: Adapted from captum llama 2 example
+# see https://captum.ai/tutorials/Llama2_LLM_Attribution
 def gpu_loading_config(max_memory: str = "15000MB"):
     n_gpus = torch.cuda.device_count()