Upload folder using huggingface_hub

grader.py

@@ -20,7 +20,7 @@ class TaskGrade:
     q_delta: float
 
 
-OPEN_INTERVAL_EPS = 1e-2
+OPEN_INTERVAL_EPS = 5e-2
 
 
 def _clamp01(x: float) -> float:

guidelines.txt

@@ -0,0 +1,113 @@
+🚀 Hackathon Submission Guidelines (OpenEnv RL Challenge)
+1. Project Structure
+Your inference script must be named inference.py
+It must be located in the root directory of your project
+
+
+
+2. LLM Usage Requirements
+You must use the OpenAI Client for all LLM calls
+Do not use alternative SDKs or direct HTTP calls
+
+
+
+3. Required Environment Variables
+Your inference.py must read the following environment variables:
+API_BASE_URL
+Description: API endpoint for the LLM
+Requirement: Must include a default value
+MODEL_NAME
+Description: Model identifier used for inference
+Requirement: Must include a default value
+HF_TOKEN
+Description: Hugging Face API token
+Requirement: Mandatory (no default required)
+
+4. INFERENCE OUTPUT FORMAT
+The script must emit exactly three line types to stdout, in this order:
+    [START] task=<task_name> env=<benchmark> model=<model_name>
+    [STEP]  step=<n> action=<action_str> reward=<0.00> done=<true|false> error=<msg|null>
+    [END]   success=<true|false> steps=<n> rewards=<r1,r2,...,rn>
+
+
+
+  Rules:
+    - One [START] line at episode begin.
+    - One [STEP] line per step, immediately after env.step() returns.
+    - One [END] line after env.close(), always emitted (even on exception).
+    - reward and rewards are formatted to 2 decimal places.
+    - done and success are lowercase booleans: true or false.
+    - error is the raw last_action_error string, or null if none.
+    - All fields on a single line with no newlines within a line.
+  Example:
+    [START] task=click-test env=miniwob model=Qwen3-VL-30B
+    [STEP] step=1 action=click('123') reward=0.00 done=false error=null
+    [STEP] step=2 action=fill('456','text') reward=0.00 done=false error=null
+    [STEP] step=3 action=click('789') reward=1.00 done=true error=null
+    [END] success=true steps=3 rewards=0.00,0.00,1.00
+✅ Example (inference.py)
+import os
+from openai import OpenAI
+
+# Read environment variables with defaults where required
+API_BASE_URL = os.getenv("API_BASE_URL", "https://api.openai.com/v1")
+MODEL_NAME = os.getenv("MODEL_NAME", "gpt-4.1-mini")
+HF_TOKEN = os.getenv("HF_TOKEN")
+
+if HF_TOKEN is None:
+    raise ValueError("HF_TOKEN environment variable is required")
+
+# Initialize OpenAI client
+client = OpenAI(
+    base_url=API_BASE_URL,
+    api_key=HF_TOKEN
+)
+
+def run_inference(prompt: str):
+    response = client.chat.completions.create(
+        model=MODEL_NAME,
+        messages=[
+            {"role": "user", "content": prompt}
+        ]
+    )
+    response = response.choices[0].message.content
+    # Print output based on above given format
+
+
+if __name__ == "__main__":
+    print(run_inference("Hello from OpenEnv!"))
+
+4. Hugging Face Space Guidelines
+Building a Hugging Face Space can take significant time, especially if multiple spaces are active
+To avoid delays:
+Turn off all unnecessary spaces
+Keep only your primary submission space running
+
+5. Submission Validation Rules
+The system will check if your Hugging Face Space is live
+If your space is not in a running state, your submission will fail automatically
+Before submitting:
+Ensure your space is fully built
+Confirm it is in the “Running” state
+
+6. Hardware Requirements
+Your solution will be executed inside a Docker container with limited resources
+It must run within the following constraints:
+2 vCPU
+8 GB RAM
+👉 Ensure your model, dependencies, and runtime fit within these limits. Submissions exceeding these constraints may fail during evaluation.
+
+6. Resubmissions
+You are allowed to resubmit your project multiple times
+If your submission fails validation, you can:
+Fix the issues
+Ensure your Hugging Face Space is running
+Submit again
+👉 There is no penalty for resubmitting, so iterate until your submission passes all checks.
+
+⚠️ Common Failure Cases (Avoid These)
+inference.py not in root directory
+Missing default values for API_BASE_URL or MODEL_NAME
+Missing HF_TOKEN
+Hugging Face Space still building during submission
+Space stopped due to multiple active deployments

inference.py

@@ -11,6 +11,7 @@ STDOUT FORMAT
 import argparse
 import asyncio
 import json
+import math
 import os
 import sys
 from pathlib import Path
@@ -39,6 +40,8 @@ BENCHMARK = os.getenv("LAYOUT_BENCHMARK", "layoutenv")
 TEMPERATURE = float(os.getenv("TEMPERATURE", "0.0"))
 MAX_TOKENS = 200
 SUCCESS_Q_DELTA = 0.1
+# Clamp epsilon — keep printed rewards strictly inside (0, 1).
+_REWARD_EPS = 0.05
 PRINT_SUMMARY_STDERR = os.getenv("PRINT_SUMMARY_STDERR", "0") == "1"
 EARLY_STOP_ON_SUCCESS = os.getenv("EARLY_STOP_ON_SUCCESS", "1") == "1"
 
@@ -67,21 +70,30 @@ def resolve_background_image_path(sample: Dict[str, Any], dataset_json_path: str
     return abs_path
 
 
+def _clamp_reward(r: float) -> float:
+    """Ensure reward is strictly inside (0, 1) — never exact 0.0 or 1.0."""
+    if r is None or math.isnan(r) or math.isinf(r):
+        return 0.5
+    return min(max(float(r), _REWARD_EPS), 1.0 - _REWARD_EPS)
+
+
 def log_start(task: str, env: str, model: str) -> None:
     print(f"[START] task={task} env={env} model={model}", flush=True)
 
 
 def log_step(step: int, action_str: str, reward: float, done: bool, error: Optional[str]) -> None:
     error_val = error if error else "null"
+    safe_reward = _clamp_reward(reward)
     print(
-        f"[STEP] step={step} action={action_str} reward={reward:.2f} "
+        f"[STEP] step={step} action={action_str} reward={safe_reward:.2f} "
         f"done={str(done).lower()} error={error_val}",
         flush=True,
     )
 
 
 def log_end(success: bool, steps: int, rewards: List[float]) -> None:
-    rewards_str = ",".join(f"{r:.2f}" for r in rewards)
+    safe_rewards = [_clamp_reward(r) for r in rewards]
+    rewards_str = ",".join(f"{r:.2f}" for r in safe_rewards)
     print(
         f"[END] success={str(success).lower()} steps={steps} rewards={rewards_str}",
         flush=True,
@@ -208,7 +220,7 @@ async def run_episode(
             action_str = action_to_string(action, raw)
             result = await env.step(action)
             obs = result.observation
-            reward = result.reward or 0.0
+            reward = _clamp_reward(result.reward)
             done = result.done
             raw_error = getattr(result, "last_action_error", None)
             if raw_error is None:

server/layout_environment.py

@@ -503,7 +503,7 @@ TERMINAL_BONUS_SCALE = 5.0
 TERMINAL_PENALTY = -1.0
 # Align terminal shaping with the easiest grader delta threshold.
 Q_DELTA_THRESHOLD = 0.05
-VISIBLE_REWARD_EPS = 0.01
+VISIBLE_REWARD_EPS = 0.05
 
 
 def _normalize_visible_reward(raw_reward: float | int) -> float: