Upload folder using huggingface_hub

.env.example

@@ -10,7 +10,7 @@
 API_BASE_URL=https://router.huggingface.co/v1
 
 # Model identifier — must be accessible at the above endpoint
-MODEL_NAME=Qwen/Qwen2.5-72B-Instruct
+MODEL_NAME=Qwen/Qwen2.5-7B-Instruct
 
 # HuggingFace API token (also used as the OpenAI-client api_key)
 HF_TOKEN=hf_your_token_here

client.py

@@ -51,18 +51,15 @@ class NL2SQLEnv:
         return self._parse_result(resp.json())
 
     def _parse_result(self, payload: Dict[str, Any]) -> StepResult:
-        obs_data = payload.get("observation", {})
+        obs_data = payload.get("observation", payload)
         
-        # SAFETY CHECK: Handle JSON 'null' (None) values gracefully
-        raw_reward = payload.get("reward")
-        safe_reward = float(raw_reward) if raw_reward is not None else 0.0
-        
-        raw_obs_reward = obs_data.get("reward")
-        safe_obs_reward = float(raw_obs_reward) if raw_obs_reward is not None else 0.0
-        
-        raw_score = obs_data.get("score")
-        safe_score = float(raw_score) if raw_score is not None else 0.0
+        # ── THE BULLETPROOF REWARD EXTRACTOR ──
+        # Check both the top-level payload and the nested observation dict.
+        val1 = float(payload.get("reward") or 0.0)
+        val2 = float(obs_data.get("reward") or 0.0)
+        safe_reward = max(val1, val2)
         
+        safe_score = float(obs_data.get("score") or 0.0)
         safe_done = bool(payload.get("done") or obs_data.get("done") or False)
 
         obs = NL2SQLObservation(
@@ -76,7 +73,7 @@ class NL2SQLEnv:
             step=obs_data.get("step", 0),
             max_steps=obs_data.get("max_steps", 5),
             done=safe_done,
-            reward=safe_obs_reward,
+            reward=safe_reward,
             score=safe_score,
         )
         return StepResult(

inference.py

@@ -27,18 +27,36 @@ from typing import List, Optional
 
 from openai import OpenAI
 
+# # ── Configuration ──────────────────────────────────────────────────────────
+# API_BASE_URL   = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
+# MODEL_NAME     = os.getenv("MODEL_NAME",   "Qwen/Qwen2.5-7B-Instruct")
+# API_KEY        = os.getenv("HF_TOKEN") or os.getenv("OPENAI_API_KEY", "")
+# IMAGE_NAME     = os.getenv("LOCAL_IMAGE_NAME", "nl2sql-bench:latest")
+# SPACE_URL      = os.getenv("SPACE_URL", "http://localhost:8000")
+
+# BENCHMARK      = "nl2sql-bench"
+# MAX_STEPS      = 5
+# TEMPERATURE    = 0.2      # Low temp for SQL generation
+# MAX_TOKENS     = 512
+# SUCCESS_THRESHOLD = 0.7   # score >= 0.7 → success
+
+# TASKS = ["simple-filter", "join-aggregation", "analytics-window"]
+
 # ── Configuration ──────────────────────────────────────────────────────────
-API_BASE_URL   = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
-MODEL_NAME     = os.getenv("MODEL_NAME",   "Qwen/Qwen2.5-7B-Instruct")
-API_KEY        = os.getenv("HF_TOKEN") or os.getenv("OPENAI_API_KEY", "")
-IMAGE_NAME     = os.getenv("LOCAL_IMAGE_NAME", "nl2sql-bench:latest")
-SPACE_URL      = os.getenv("SPACE_URL", "http://localhost:8000")
-
-BENCHMARK      = "nl2sql-bench"
-MAX_STEPS      = 5
-TEMPERATURE    = 0.2      # Low temp for SQL generation
-MAX_TOKENS     = 512
-SUCCESS_THRESHOLD = 0.7   # score >= 0.7 → success
+API_BASE_URL      = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
+# Points to your newly uploaded fine-tuned weights!
+MODEL_NAME        = os.getenv("MODEL_NAME", "ritvik360/qwen-7b-nl2sql-merged_1") 
+# CRITICAL FIX: Looks for 'API_KEY' first to satisfy the evaluator's LiteLLM proxy
+API_KEY           = os.getenv("API_KEY")  or os.getenv("HF_TOKEN", "") or os.getenv("OPENAI_API_KEY")
+IMAGE_NAME        = os.getenv("LOCAL_IMAGE_NAME", "nl2sql-bench:latest")
+# CRITICAL FIX: Point the default directly to your live HF Space!
+SPACE_URL         = os.getenv("SPACE_URL", "https://ritvik360-nl2sql-bench.hf.space")
+
+BENCHMARK         = "nl2sql-bench"
+MAX_STEPS         = 5
+TEMPERATURE       = 0.2      # Low temp for SQL generation
+MAX_TOKENS        = 512
+SUCCESS_THRESHOLD = 0.7      # score >= 0.7 → success
 
 TASKS = ["simple-filter", "join-aggregation", "analytics-window"]