Update app.py

app.py

@@ -4,64 +4,50 @@ import json
 import os
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from peft import PeftModel
-from huggingface_hub import InferenceClient, snapshot_download
+from huggingface_hub import InferenceClient
 
 # ==============================================================================
 # 1. CONFIGURATION
 # ==============================================================================
-HF_TOKEN = os.getenv("HF_TOKEN")
+# NOTE: You must set 'HF_TOKEN' in your Hugging Face Space Secrets!
+HF_TOKEN = os.getenv("HF_TOKEN") 
+
 PROJECT_TITLE = "The Janus Interface: Semantic Decoupling Architecture"
 
 # Models
+# We use the official Microsoft repo for CPU compatibility
 BASE_MODEL_ID = "microsoft/Phi-3.5-mini-instruct" 
 ADAPTER_ID = "st192011/janus-gold-lora" 
 CLOUD_MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct" 
 
 # ==============================================================================
-# 2. ENGINE INITIALIZATION (CPU Optimized + Config Sanitizer)
+# 2. ENGINE INITIALIZATION (CPU Optimized)
 # ==============================================================================
-print("⏳ Initializing Neural Backbone...")
+print("⏳ Initializing Neural Backbone (CPU Mode)...")
 
 try:
+    # Load Tokenizer
     tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
     
-    # Load Base Model
+    # Load Base Model (bfloat16 saves RAM on Free Tier Spaces)
     base_model = AutoModelForCausalLM.from_pretrained(
         BASE_MODEL_ID,
         torch_dtype=torch.bfloat16, 
         device_map="cpu",
-        trust_remote_code=True,
-        attn_implementation="eager" 
+        trust_remote_code=True
     )
     
-    print(f"⏳ Downloading and sanitizing adapter ({ADAPTER_ID})...")
-    
-    # 1. Download the adapter files locally
-    local_adapter_path = snapshot_download(repo_id=ADAPTER_ID, token=HF_TOKEN)
-    
-    # 2. Load the config JSON
-    config_path = os.path.join(local_adapter_path, "adapter_config.json")
-    with open(config_path, "r") as f:
-        config_data = json.load(f)
-    
-    # 3. Remove the key that causes the crash
-    if "alora_invocation_tokens" in config_data:
-        print("🧹 Cleaning incompatible Unsloth config keys...")
-        del config_data["alora_invocation_tokens"]
-        
-        # Save the clean config back to disk
-        with open(config_path, "w") as f:
-            json.dump(config_data, f, indent=2)
-            
-    # 4. Load the adapter from the local sanitized folder
-    model = PeftModel.from_pretrained(base_model, local_adapter_path)
-    model.eval()
+    # Load Adapter
+    print(f"⏳ Mounting Janus Adapter ({ADAPTER_ID})...")
+    model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
+    model.eval() # Set to inference mode
     print("✅ System Online.")
 
 except Exception as e:
     print(f"❌ Error loading model: {e}")
     raise e
 
+# Cloud Client
 hf_client = InferenceClient(model=CLOUD_MODEL_ID, token=HF_TOKEN)
 
 # ==============================================================================
@@ -70,11 +56,15 @@ hf_client = InferenceClient(model=CLOUD_MODEL_ID, token=HF_TOKEN)
 
 def clean_output(text):
     """Sanitizes output to prevent chain-reaction failures."""
+    # Remove special tokens
     clean = text.replace("<|end|>", "").replace("<|endoftext|>", "")
+    
+    # Remove conversational filler lines
     if "Output:" in clean: clean = clean.split("Output:")[-1]
     
     lines = clean.split('\n')
-    valid_lines = [line for line in lines if ":" in line and "Note" not in line]
+    # Keep lines that look like protocol code or normal text, remove "Here is..."
+    valid_lines = [line for line in lines if "Note" not in line and "Here is" not in line]
     return " ".join(valid_lines).strip()
 
 def kernel_scout(raw_input):
@@ -100,7 +90,7 @@ RAW NOTE:
                 max_new_tokens=256, 
                 temperature=0.1,
                 do_sample=True,
-                use_cache=True
+                use_cache=False
             )
         
         text = tokenizer.batch_decode(outputs)[0]
@@ -154,11 +144,11 @@ PRIVATE_DB:
         with torch.no_grad():
             outputs = model.generate(
                 **inputs, 
-                max_new_tokens=600, 
+                max_new_tokens=1024, 
                 temperature=0.1, 
                 repetition_penalty=1.05,
                 do_sample=True,
-                use_cache=True
+                use_cache=False
             )
         
         text = tokenizer.batch_decode(outputs)[0]
@@ -213,6 +203,8 @@ report_md = """
 # 🏛️ The Janus Interface: Research & Technical Analysis
 **Project Status:** Research Prototype v2.0 (Gold Standard)
 
+---
+
 ### 1. Research Motivation: The Privacy-Utility Paradox
 In regulated domains (Healthcare, Legal, Finance), Generative AI adoption is stalled by a fundamental conflict:
 *   **Utility:** Large Cloud Models (GPT-4, Claude) offer superior reasoning but require sending data off-premise.
@@ -235,12 +227,14 @@ The system utilizes a **Multi-Task Adapter** trained to switch between two disti
 *   **Function:** A secure, offline engine that accepts the JanusScript and a Local SQL Database record.
 *   **Output:** It merges the abstract logic with the concrete identity to generate the final, human-readable document.
 
+---
+
 ### 3. Data Engineering: The "Gold Standard" Pipeline
-To achieve high fidelity without using private patient data, we developed a **Teacher-Student Distillation** pipeline:
-1.  **Source:** **MTSamples** (Open Source Medical Transcription).
-2.  **Distillation:** We used **Llama-3-70B** to read 4,000+ real medical notes and extract the logic into our custom `JanusScript` syntax.
-3.  **Synthesis:** We generated synthetic identities (Names/MRNs) using Python libraries.
-4.  **Alignment:** We programmatically constructed the "Target Output" by prepending the synthetic header to the real medical text.
+To achieve high fidelity without using private patient data, we developed a **Synthesized Data Pipeline**:
+
+1.  **Synthesis:** We generated **306 high-quality clinical scenarios** using Large Language Models (LLMs).
+2.  **Alignment:** Unlike previous iterations where headers were random, this dataset ensured strict mathematical alignment between the Identity Header (Age/Sex) and the Clinical Narrative.
+3.  **Result:** This eliminated the "hallucination" issues seen in earlier tests where the model would confuse patient gender or age due to conflicting training signals.
 
 ### 4. Training Methodology
 *   **Base Model:** Microsoft Phi-3.5-mini-instruct (3.8B Parameters).
@@ -248,10 +242,10 @@ To achieve high fidelity without using private patient data, we developed a **Te
 *   **Technique:** **DoRA (Weight-Decomposed Low-Rank Adaptation)**.
     *   *Why DoRA?* Standard LoRA struggles with strict syntax/coding tasks. DoRA updates both magnitude and direction vectors, allowing the model to learn the strict `JanusScript` grammar effectively.
 *   **Loss Masking:** We used `train_on_responses_only`. The model was **never** trained on the input text, only on the output. This prevents the model from memorizing patient PII from the training set.
-*   **Hyperparameters:** Rank 16, Alpha 16, Learning Rate 2e-4, 2 Epochs (306 samples).
+*   **Hyperparameters:** Rank 16, Alpha 16, Learning Rate 2e-4, **2 Epochs** (approx 78 steps used for final checkpoint).
 
 ### 5. Results & Conclusion
-*   **Zero-Trust Validation:** The "Vault" successfully reconstructs documents using *only* the database for identity. If the DB says "Male" and the training data said "Female," the model now correctly obeys the DB.
+*   **Zero-Trust Validation:** The "Vault" successfully reconstructs documents using *only* the database for identity.
 *   **Semantic Expansion:** The model demonstrates the ability to take a concise code (`Dx(Pneumonia)`) and expand it into fluent medical narrative ("Patient presented with symptoms consistent with Pneumonia...").
 """