Upload 3 files

config_physics.py

@@ -1,41 +1,49 @@
-
-import torch
-
-class Config:
-    # Model
-    MODEL_ID = "unsloth/gemma-3-1b-it" 
-    # MODEL_ID = "google/gemma-3-1b-it" 
-    
-    # Dimensions
-    HIDDEN_SIZE = 1152 # Gemma 3 1B hidden size
-    LATENT_DIM = 256   # Physics latent state dimension
-    PROJECTOR_HIDDEN = 1024
-    
-    # Physics Controller
-    CONTROLLER_HIDDEN = 512
-    MODULATION_DIM = 64 # Rank of modulation (similar to LoRA rank)
-    
-    # Training (OPTIMIZED for Contrastive Physics)
-    BATCH_SIZE = 1          # Back to 1: Contrastive needs 4 forward passes!
-    GRAD_ACCUMULATION = 64  # 1 * 64 = 64 effective batch
-    LEARNING_RATE = 5e-5    # Reduced: Controller needs stability with Lazy Tax
-    POLICY_LR = 5e-5        # Sync with LR
-    EPOCHS = 3              # Reduced: Contrastive Loss accelerates learning
-    
-    # TTT / Inference
-    TTT_STEPS = 5
-    TTT_LR = 1e-4
-    
-    # Data
-    MAX_LENGTH = 256 # Reduced: Math reasoning is shorter. Saves VRAM.
-    DTYPE = torch.float32
-    
-    # Physics Dimensions
-    PHYSICS_DIMS = [
-        "Gravity", "Friction", "Elasticity", "Fragility", "Density", 
-        "Temperature", "Conductivity", "Magnetism",
-        "Thermodynamics", "Fluid Dynamics", "Vacuum Physics", "Electromagnetism"
-    ]
-    
-    # Keys
-    API_KEY_ENV = "GEMINI_API_KEY"
+
+import torch
+
+class Config:
+    # Model
+    MODEL_ID = "unsloth/gemma-3-1b-it" 
+    # MODEL_ID = "google/gemma-3-1b-it" 
+    
+    # Dimensions
+    HIDDEN_SIZE = 1152 # Gemma 3 1B hidden size
+    LATENT_DIM = 256   # Physics latent state dimension
+    PROJECTOR_HIDDEN = 1024
+    
+    # Physics Controller
+    CONTROLLER_HIDDEN = 512
+    MODULATION_DIM = 64 # Rank of modulation (similar to LoRA rank)
+    
+    # Training (OPTIMIZED for Contrastive Physics)
+    BATCH_SIZE = 1          # Back to 1: Contrastive needs 4 forward passes!
+    GRAD_ACCUMULATION = 64  # 1 * 64 = 64 effective batch
+    LEARNING_RATE = 5e-5    # Reduced: Controller needs stability with Lazy Tax
+    POLICY_LR = 5e-5        # Sync with LR
+    EPOCHS = 3              # Reduced: Contrastive Loss accelerates learning
+    
+    # TTT / Inference
+    TTT_STEPS = 5
+    TTT_LR = 1e-4
+    
+    # Data
+    MAX_LENGTH = 256 # Reduced: Math reasoning is shorter. Saves VRAM.
+    DTYPE = torch.float32
+    
+    # Physics Dimensions
+    PHYSICS_DIMS = [
+        "Gravity", "Friction", "Elasticity", "Fragility", "Density", 
+        "Temperature", "Conductivity", "Magnetism",
+        "Thermodynamics", "Fluid Dynamics", "Vacuum Physics", "Electromagnetism"
+    ]
+    
+    # Keys
+    API_KEY_ENV = "GEMINI_API_KEY"
+    
+    # Text Generation / Prompting
+    SYSTEM_PROMPT = (
+        "You are Flux, an advanced physics simulation engine. "
+        "You answer questions based on precise physical laws, distinguishing between environments like "
+        "Vacuum, Earth, Moon, and Zero-G. "
+        "Think step-by-step using first principles (Newton's Laws, Gravity, Fluid Dynamics)."
+    )

continuous_learning_cumulative.py

@@ -82,6 +82,18 @@ def run_cumulative_ttt():
     
     # 3. Curriculum
     curriculum = [
+        {
+            "id": "scenario_general",
+            "concept": "General Language Grounding",
+            "prompt": "What is the capital of France?",
+            "correction": "The capital of France is Paris.",
+            "test_variations": [
+                {"q": "Summarize: The quick brown fox jumps over the lazy dog.", "a": "A fox jumps over a dog."},
+                {"q": "What is 2 + 2?", "a": "4"},
+                {"q": "Explain what a tree is.", "a": "A tree is a tall plant with a trunk and branches made of wood."},
+                {"q": "Who wrote Romeo and Juliet?", "a": "William Shakespeare."}
+            ]
+        },
         {
             "id": "scenario_1",
             "concept": "Zero Gravity Inertia",
@@ -142,6 +154,11 @@ def run_cumulative_ttt():
             replay_buffer.add(task['id'], v['q'], v['a'])
         replay_buffer.add(task['id'], task['prompt'], task['correction'])
         
+        # SKIP TRAINING for General Scenario (Just use as Replay Anchor)
+        if task['id'] == "scenario_general":
+            print("   ⏩ Added General Anchors to Buffer. Skipping Training Step.")
+            continue
+            
         print("   🧠 Robust Learning (Current + Stratified History)...")
         model.train()
         
@@ -204,7 +221,7 @@ def run_cumulative_ttt():
             q = item['q']
             target = item['a']
             print(f"   Q: \"{q}\"")
-            inputs = model.tokenizer(f"User: {q}\nModel:", return_tensors="pt").to(device)
+            inputs = model.tokenizer(f"{Config.SYSTEM_PROMPT}\nUser: {q}\nModel:", return_tensors="pt").to(device)
             with torch.no_grad():
                  h_init = model.get_embeddings(inputs.input_ids).to(Config.DTYPE)
                  mod = model.controller(h_init)
@@ -250,7 +267,7 @@ def run_cumulative_ttt():
     print("✅ Model Saved: final_physics_controller.pt, final_flux_adapters.pt")
 
 def calculate_loss(model, prompt, answer, device):
-    full_text = f"User: {prompt}\nModel: {answer}"
+    full_text = f"{Config.SYSTEM_PROMPT}\nUser: {prompt}\nModel: {answer}"
     inputs = model.tokenizer(full_text, return_tensors="pt").to(device)
     
     # Forward Pass
@@ -269,6 +286,7 @@ def check_answer(task_id, text):
     if task_id == "scenario_2": return "same time" in text or "equal" in text or "identical" in text or "same rate" in text or "neither" in text or "instant" in text or "side-by-side" in text
     if task_id == "scenario_3": return "up" in text or "rise" in text or "float" in text
     if task_id == "scenario_4": return "coin" in text or "thrown" in text or "initial" in text or "toss" in text or "rock" in text or "bullet" in text or "volleyball" in text
+    if task_id == "scenario_general": return "paris" in text or "fox" in text or "4" in text or "plant" in text or "shakespeare" in text
     return False
 
 if __name__ == "__main__":

inference_physics.py

@@ -65,8 +65,8 @@ def interactive_session():
             if not user_input.strip():
                 continue
                 
-            # Format prompt EXACTLY like training
-            full_prompt = f"User: {user_input}\nModel:"
+            # Format prompt EXACTLY like training (System Prompt + Chat)
+            full_prompt = f"{Config.SYSTEM_PROMPT}\nUser: {user_input}\nModel:"
             
             inputs = model.tokenizer(full_prompt, return_tensors="pt").to(device)