refactor: remove post-processing constraints in app.py and update model training to export HF merged format via llama.cpp conversion

app.py

@@ -64,6 +64,8 @@ async def run_blocking(func, *args, **kwargs):
 async def generate_atomic_task(goal: str, previous_failures: int, history: list = None) -> str:
     if history is None:
         history = []
+    history_str = "\n".join([f"- {t}" for t in history]) if history else "None"
+    
     if MOCK_MODE:
         await asyncio.sleep(1) # Simulate inference latency
         demo_steps = [
@@ -108,14 +110,6 @@ async def generate_atomic_task(goal: str, previous_failures: int, history: list
             
         content = content.strip('"\' ,.-1234567890)')
         
-        # POST-PROCESSING: 4B models often ignore constraints and try to output lists.
-        # Force it to be atomic by taking the first item before a comma or " and ".
-        if "," in content:
-            content = content.split(",")[0].strip()
-        if " and " in content.lower():
-            content = content.split(" and ")[0].strip()
-            
-        content = content.strip('"\' ,.-1234567890)')
         
         print(f"NEMOTRON RAW OUT: {content}", flush=True)
         return content or "Focus on the screen."

modal_train.py

@@ -106,27 +106,26 @@ def train_model():
     )
     trainer.train()
     
-    print("Exporting to GGUF...")
-    import os
-    os.makedirs("export_dir", exist_ok=True)
-    model.save_pretrained_gguf("export_dir", tokenizer, quantization_method = "q4_k_m")
+    print("Exporting HuggingFace 16-bit Model...")
+    model.save_pretrained_merged("step-zero-nemotron-hf", tokenizer, save_method="merged_16bit")
     
-    import shutil
-    import glob
-    import os
+    import subprocess
+    print("Cloning llama.cpp to run conversion manually...")
+    subprocess.run(["git", "clone", "https://github.com/ggerganov/llama.cpp.git"])
+    subprocess.run(["pip", "install", "-r", "llama.cpp/requirements.txt"])
     
-    print("Files in export_dir:", os.listdir("export_dir"))
+    print("Converting to GGUF using llama.cpp...")
+    subprocess.run([
+        "python3", "llama.cpp/convert_hf_to_gguf.py", 
+        "step-zero-nemotron-hf", 
+        "--outfile", "/vol/step-zero-nemotron-finetuned.gguf", 
+        "--outtype", "q8_0"
+    ])
     
-    gguf_files = glob.glob("export_dir/*.gguf") + glob.glob("*.gguf")
-    for file in gguf_files:
-        print(f"Copying {file} to /vol...")
-        shutil.copy(file, "/vol/step-zero-nemotron.gguf")
-        
     try:
         vol.commit()
     except Exception as e:
         print("vol.commit() skipped:", e)
-        
     print("Training complete. GGUF artifact generated in /vol.")
 
 @app.local_entrypoint()