Update app.py

app.py

@@ -1,35 +1,141 @@
 import gradio as gr
-from transformers import pipeline
+from llama_cpp import Llama
+import os
+
+# Define the model path. Hugging Face's model hub automatically handles
+# downloading the specified GGUF file.
+# We use the fp16 version as it's commonly available and works well with GGUF.
+# You can explore other quantized versions like Q4_K_M if needed, but fp16 is a safe start.
+model_name = "microsoft/Phi-3-mini-4k-instruct-gguf"
+model_file = "phi-3-mini-4k-instruct-fp16.gguf"
+model_path = hf_hub_download(repo_id=model_name, filename=model_file)
+
+# Load the GGUF model using llama-cpp-python
+# n_gpu_layers=0 forces the model to run on the CPU, which is required for free spaces.
+# n_ctx sets the context length, 4096 is appropriate for this model version.
+# verbose=False reduces logging output.
+try:
+    llm = Llama(model_path=model_path, n_gpu_layers=0, n_ctx=4096, verbose=False)
+    print("Model loaded successfully on CPU.")
+except Exception as e:
+    print(f"Error loading model: {e}")
+    # Handle the error - maybe raise it or set llm to None
+    llm = None
 
-classifier = pipeline("zero-shot-classification", model="lytang/MiniCheck-Flan-T5-Large")
-labels = ["True", "False", "Unknown"]
 
 def fact_check(statement):
-    result = classifier(statement, candidate_labels=labels)
-    label = result['labels'][0]
-    confidence = result['scores'][0]
-
-    explanation = (
-        f"The model classified the statement as **{label}** "
-        f"with a confidence of {confidence:.2f}. "
-    )
-    
-    if label == "True":
-        explanation += "This suggests the statement aligns with commonly known facts."
-    elif label == "False":
-        explanation += "This suggests the statement contradicts widely accepted information."
-    else:
-        explanation += "The model is unsure due to lack of context or conflicting evidence."
-
-    return explanation
-
-with gr.Blocks(title="AI Fact Checker") as demo:
-    gr.Markdown("### Enter a statement. The model will check if it's True, False, or Unknown.")
-    input_text = gr.Textbox(lines=3, label="Statement")
-    output_md = gr.Markdown(label="Fact Check & Reasoning")
+    if llm is None:
+        return "Error: Language model failed to load."
+
+    # Craft a prompt for the instruction-tuned model.
+    # We instruct the model to classify and provide a reason in a specific format
+    # that we can then parse.
+    prompt = f"""<|user|>
+Assess the truthfulness of the following statement.
+Classify it as True, False, Unknown, or Misleading.
+Provide a brief reason for your classification based on general knowledge.
+
+Statement: {statement}
+
+Provide the output in the following format:
+Classification: [Your Classification]
+Reason: [Your Reason]
+<|end|>
+<|assistant|>
+Classification:""" # Start the assistant's response to guide the generation
+
+    try:
+        # Generate the completion from the model
+        # max_tokens: Maximum number of tokens to generate for the response.
+        # stop: Stop generation when these tokens are encountered. <|end|> is the EOS token for Phi-3.
+        # temperature: Controls randomness (lower means more deterministic).
+        # top_p: Controls diversity of output (lower means more focused).
+        output = llm.create_completion(
+            prompt,
+            max_tokens=256, # Adjust as needed for longer reasons
+            stop=["<|end|>"],
+            temperature=0.1,
+            top_p=0.9,
+            repeat_penalty=1.1, # Helps prevent repetitive text
+        )
+
+        # Extract the generated text from the model's output
+        generated_text = output['choices'][0]['text'].strip()
+
+        # The model's response starts after "Classification:".
+        # We need to prepend "Classification:" to the generated text for parsing.
+        full_output_text = "Classification:" + generated_text
+
+        # Parse the output to extract the classification and reason
+        lines = full_output_text.split('\n')
+        label = "Unknown" # Default if parsing fails
+        reason = "Could not extract a reason from the model's response." # Default reason
+
+        extracted_label = "Unknown"
+        extracted_reason = ""
+
+        # Simple line-by-line parsing based on the expected format
+        for line in lines:
+            if line.startswith("Classification:"):
+                extracted_label = line.replace("Classification:", "").strip()
+            elif line.startswith("Reason:"):
+                extracted_reason = line.replace("Reason:", "").strip()
+
+        # Map the extracted label to the desired categories, handling variations
+        lower_label = extracted_label.lower()
+        if 'true' in lower_label:
+            classification = 'True'
+        elif 'false' in lower_label:
+            classification = 'False'
+        elif 'misleading' in lower_label:
+            classification = 'Misleading'
+        else:
+            classification = 'Unknown' # Catches 'Unknown' and anything else
+
+
+        # Use the extracted reason, or a default if it's empty after extraction
+        reason = extracted_reason if extracted_reason else "No specific reason provided by the model."
+
+        # Construct the final explanation markdown
+        explanation_md = f"**Statement:** {statement}\n\n"
+        explanation_md += f"**Classification:** <span style='color: {'green' if classification == 'True' else 'red' if classification == 'False' else 'orange'}'>{classification}</span>\n\n"
+        explanation_md += f"**Reason:** {reason}"
+
+    except Exception as e:
+        # Catch any errors during the generation or parsing process
+        explanation_md = f"An error occurred during fact checking: {e}"
+        classification = "Error" # Indicate an error state
+
+    return explanation_md
+
+# Gradio Interface Definition
+with gr.Blocks(title="Lightweight AI Fact Checker (CPU)") as demo:
+    gr.Markdown("""
+    ### Lightweight AI Fact Checker on Hugging Face Spaces (CPU)
+    Enter a statement below. A lightweight model (Phi-3 Mini GGUF) running on a free CPU tier will attempt to classify its truthfulness (True, False, Unknown, or Misleading) and provide a brief reason.
+    """)
+
+    input_text = gr.Textbox(lines=3, label="Statement to Fact Check")
+    output_md = gr.Markdown(label="Fact Check Results")
+
     btn = gr.Button("Check Fact")
-    
+
+    # Set the function to call when the button is clicked
     btn.click(fn=fact_check, inputs=input_text, outputs=output_md)
 
+# Launch the Gradio app
+# server_name="0.0.0.0" and server_port=7860 are necessary for Hugging Face Spaces
 if __name__ == "__main__":
-    demo.launch()
+    # Add a download step for the model file outside the function
+    # This ensures it's downloaded when the app starts in the Space
+    try:
+        from huggingface_hub import hf_hub_download
+        print(f"Downloading model: {model_name}/{model_file}")
+        hf_hub_download(repo_id=model_name, filename=model_file)
+        print("Download complete.")
+    except Exception as e:
+        print(f"Error during model download: {e}")
+        # The model loading later will also catch this, but this gives an earlier error message.
+
+    print("Launching Gradio app...")
+    demo.launch(server_name="0.0.0.0", server_port=7860)