Fix merge: fall back to warm-start adapter from HF when GRPO skipped

train_on_hf.py

@@ -331,16 +331,29 @@ def train_strategist(data_dir: Path, max_samples: int = 10000):
 
 def merge_and_push(hf_token: str):
     """Merge LoRA, push merged model to HF Hub."""
+    import os
     from transformers import AutoModelForCausalLM, AutoTokenizer
     from peft import PeftModel
 
     config = json.load(open("data/preprocessing_config.json"))
     MODEL_NAME = config["model"]["name"]
 
-    print("\n=== Merging LoRA and pushing to HF ===")
+    # Use strategist_final if it exists, otherwise fall back to warm-start
+    adapter_path = "./strategist_final" if os.path.exists("./strategist_final/adapter_config.json") else "./strategist_warmstart"
+    if not os.path.exists(adapter_path):
+        # If neither local dir exists, download the warm-start from HF
+        from huggingface_hub import snapshot_download
+        adapter_path = snapshot_download(
+            repo_id="Rayugacodes/kernelx-strategist",
+            allow_patterns=["adapter/*"],
+            local_dir="./hf_adapter",
+        )
+        adapter_path = "./hf_adapter/adapter"
+
+    print(f"\n=== Merging LoRA from {adapter_path} and pushing to HF ===")
     base = AutoModelForCausalLM.from_pretrained(MODEL_NAME, device_map="cpu")
     tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
-    model = PeftModel.from_pretrained(base, "./strategist_final")
+    model = PeftModel.from_pretrained(base, adapter_path)
     merged = model.merge_and_unload()
 
     merged.save_pretrained("./strategist_merged")