Upload folder using huggingface_hub

README.md

@@ -54,7 +54,7 @@ openenv-data-cleaning/
 | `normalize_category` | Categorical column | `{}` | Only valid when case-only category inconsistencies remain. |
 | `create_feature` | Registered feature name | `{"feature_name": "<name>"}` | Feature must be required by the task and its source column must already be clean enough to use. |
 
-Invalid actions leave the dataset unchanged, emit `{"error": "invalid_action"}` in `info`, consume a step, and return reward `-0.05`.
+Invalid actions leave the dataset unchanged, emit `{"error": "invalid_action"}` in `info`, consume a step, and return a low reward `0.01`.
 
 ## Observation and State Space
 
@@ -89,6 +89,7 @@ Step reward:
 ```text
 reward = (new_quality - old_quality) + ordering_bonus - 0.01
 ordering_bonus = 0.05 if dependencies were already satisfied else 0.0
+reward is then clamped to `(0.01, 0.99)`
 ```
 
 Dataset quality score combines:

env/graders.py

@@ -10,4 +10,10 @@ class DataCleaningGrader:
         penalty = wrong_actions * 0.05
 
         score = 0.8 * correctness + 0.2 * efficiency - penalty
-        return round(max(0.0, min(1.0, score)), 2)
+        # Phase 2 requires task scores to stay strictly inside (0, 1).
+        rounded_score= round(max(0.01, min(0.99, score)), 2)
+        if rounded_score <= 0.0:
+            return 0.01
+        if rounded_score >= 1.0:
+            return 0.99
+        return rounded_score

env/quality.py

@@ -51,7 +51,7 @@ def _compute_consistency(dataset: list[dict], column_infos: list) -> float:
 
 def compute_quality_score(dataset: list[dict], column_infos: list, original_issues_count: int) -> float:
     if original_issues_count == 0:
-        return 1.0
+        return 0.99
 
     total_cells = len(dataset) * len(dataset[0]) if dataset else 1
     missing_cells = sum(
@@ -65,4 +65,5 @@ def compute_quality_score(dataset: list[dict], column_infos: list, original_issu
 
     consistency = _compute_consistency(dataset, column_infos)
 
-    return round(0.4 * completeness + 0.3 * uniqueness + 0.3 * consistency, 4)
+    score = 0.4 * completeness + 0.3 * uniqueness + 0.3 * consistency
+    return round(max(0.01, min(0.99, score)), 4)

env/rewards.py

@@ -5,10 +5,16 @@ def compute_reward(
     resolved_dependency_correctly: bool,
 ) -> float:
     if not action_valid:
-        return -0.05
+        return 0.01
 
     progress = new_quality - old_quality
     ordering_bonus = 0.05 if resolved_dependency_correctly else 0.0
     step_cost = -0.01
 
-    return round(progress + ordering_bonus + step_cost, 4)
+    reward = progress + ordering_bonus + step_cost
+    rounded_score= round(max(0.01, min(0.99, score)), 2)
+    if rounded_score <= 0.0:
+        return 0.01
+    if rounded_score >= 1.0:
+        return 0.99
+    return rounded_score

inference.py

@@ -15,8 +15,8 @@ from env.graders import DataCleaningGrader
 from env.models import Action
 
 HF_TOKEN = os.getenv("HF_TOKEN")
-API_BASE_URL = os.getenv("API_BASE_URL")
-MODEL_NAME = os.getenv("MODEL_NAME")
+API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
+MODEL_NAME = os.getenv("MODEL_NAME", "openai/gpt-oss-120b")
 BENCHMARK = "data_cleaning_env"
 
 TASKS = ["basic_cleaning", "moderate_cleaning", "full_pipeline"]
@@ -91,10 +91,10 @@ def log_step(step, action_str, reward, done, error):
     )
 
 
-def log_end(success, steps, score, rewards):
+def log_end(success, steps, rewards):
     rewards_str = ",".join(f"{reward:.2f}" for reward in rewards)
     success_val = str(success).lower()
-    print(f"[END] success={success_val} steps={steps} score={score:.2f} rewards={rewards_str}", flush=True)
+    print(f"[END] success={success_val} steps={steps} rewards={rewards_str}", flush=True)
 
 
 def run_task(task_name: str):
@@ -126,7 +126,7 @@ def run_task(task_name: str):
             response = client.chat.completions.create(
                 model=require_env("MODEL_NAME", MODEL_NAME),
                 messages=messages,
-                temperature=0.3,
+                temperature=0.0,
                 max_tokens=200,
             )
             response_text = response.choices[0].message.content or ""
@@ -143,8 +143,8 @@ def run_task(task_name: str):
 
         except Exception as exc:
             step_count += 1
-            rewards_list.append(-0.05)
-            log_step(step_count, "parse_error", -0.05, False, str(exc))
+            rewards_list.append(0.01)
+            log_step(step_count, "parse_error", 0.01, False, str(exc))
             messages.append(
                 {
                     "role": "user",
@@ -163,7 +163,7 @@ def run_task(task_name: str):
             "max_steps": max_possible_steps,
         },
     )
-    log_end(success, step_count, task_score, rewards_list)
+    log_end(success, step_count, rewards_list)
     return task_score
 
 

openenv.yaml

@@ -15,7 +15,7 @@ action_space:
   type: dict
   description: "Action with action_type, column, and params fields"
 
-reward_range: [-0.05, 1.0]
+reward_range: [0.01, 0.99]
 
 tasks:
   - name: basic_cleaning

server/app.py

@@ -5,6 +5,7 @@ from typing import Any, Literal
 
 import uvicorn
 from fastapi import Body, FastAPI
+from fastapi.responses import RedirectResponse, Response
 from pydantic import BaseModel
 
 from models import Action, Observation
@@ -44,6 +45,21 @@ def root() -> dict[str, Any]:
     return payload
 
 
+@app.get("/web", include_in_schema=False)
+def web_root() -> RedirectResponse:
+    return RedirectResponse(url="/", status_code=307)
+
+
+@app.get("/web/", include_in_schema=False)
+def web_root_slash() -> RedirectResponse:
+    return RedirectResponse(url="/", status_code=307)
+
+
+@app.get("/favicon.ico", include_in_schema=False)
+def favicon() -> Response:
+    return Response(status_code=204)
+
+
 @app.get("/health")
 def health() -> dict[str, str]:
     return {"status": "healthy"}

test_env.py

@@ -17,7 +17,7 @@ def assert_invalid_action_consumes_step() -> None:
     _, reward, _, info = env.step(
         Action(action_type="convert_dtype", column="age", params={"target_dtype": "int"})
     )
-    assert reward == -0.05
+    assert reward == 0.01
     assert info["error"] == "invalid_action"
     assert env.steps_remaining == obs.steps_remaining - 1
 
@@ -28,7 +28,7 @@ def assert_dependency_gate() -> None:
     _, reward, _, info = env.step(
         Action(action_type="convert_dtype", column="salary", params={"target_dtype": "int"})
     )
-    assert reward == -0.05
+    assert reward == 0.01
     assert info["error"] == "invalid_action"