Fix: point PeftModel.from_pretrained at the correct adapter subfolder (rl/seed_42/best)

app.py

@@ -14,6 +14,11 @@ from peft import PeftModel
 
 BASE_MODEL = "Qwen/Qwen2.5-Coder-14B-Instruct"
 ADAPTER = "adelelsayed1991/fhirsql-reasoning-sql-adapters"
+# The adapters repo holds six checkpoints (sft/rl x seed_42/43/44), each in
+# its own subfolder rather than at the repo root. rl/seed_42/best is served
+# here as the project's most advanced trained artifact (SFT+RL); change
+# this to any of sft|rl/seed_{42,43,44}/best to serve a different one.
+ADAPTER_SUBFOLDER = "rl/seed_42/best"
 ABSTENTION_TOKEN = "UNANSWERABLE"
 MAX_NEW_TOKENS = 512
 
@@ -44,7 +49,7 @@ SYSTEM_PROMPT = SYSTEM_PROMPT_TEMPLATE.format(
 
 tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
 base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype="auto")
-model = PeftModel.from_pretrained(base_model, ADAPTER)
+model = PeftModel.from_pretrained(base_model, ADAPTER, subfolder=ADAPTER_SUBFOLDER)
 model.to("cuda")
 model.eval()