Upload folder using huggingface_hub

backend.py

@@ -105,10 +105,11 @@ class Inference:
             with torch.no_grad():
                 outputs = model.generate(
                     **inputs,
-                    max_new_tokens=256,
+                    max_new_tokens=512,
                     do_sample=True,
-                    temperature=0.7,
-                    top_p=0.9,
+                    temperature=0.4,
+                    top_p=0.85,
+                    repetition_penalty=1.15,
                     pad_token_id=tokenizer.eos_token_id,
                 )
             logger.info(f"Generated output shape: {outputs.shape}")

deploy.py

@@ -4,7 +4,6 @@
 import os
 import re
 import subprocess
-import sys
 
 from dotenv import load_dotenv
 from huggingface_hub import HfApi, SpaceHardware
@@ -14,11 +13,17 @@ load_dotenv()
 SPACE_TITLE = "posttraining-practice"
 
 
+def require_env(name: str) -> str:
+    """Get required environment variable or exit with error."""
+    value = os.environ.get(name)
+    if value is None:
+        raise SystemExit(f"ERROR: {name} must be set in .env")
+    return value
+
+
 def main():
-    api_key = os.environ.get("MODEL_SITE_API_KEY")
-    site_password = os.environ.get("SITE_PASSWORD")
-    if api_key is None or site_password is None:
-        sys.exit("ERROR: MODEL_SITE_API_KEY and SITE_PASSWORD must be set in .env")
+    api_key = require_env("MODEL_SITE_API_KEY")
+    site_password = require_env("SITE_PASSWORD")
 
     api = HfApi()
     user = api.whoami()["name"]
@@ -35,7 +40,7 @@ def main():
 
     match = re.search(r"https://[^\s]+\.modal\.run", result.stdout + result.stderr)
     if match is None:
-        sys.exit("ERROR: Could not find Modal endpoint URL")
+        raise SystemExit("ERROR: Could not find Modal endpoint URL")
     modal_endpoint = match.group(0)
 
     # Generate requirements.txt

models.py

@@ -1,6 +1,7 @@
 MODEL_IDS = [
     "sachiniyer/Qwen2.5-0.5B-DPO-Schwinn",
     "sachiniyer/Qwen2.5-0.5B-PPO-Schwinn",
+    "Qwen/Qwen2.5-0.5B",
     "sachiniyer/SmolLM2-DPO-Schwinn-SmolLM2-Base",
     "sachiniyer/SmolLM2-DPO-Schwinn-gpt-5-mini-base",
     "sachiniyer/SmolLM2-FT-SFT-Learning",