Spaces:
Sleeping
Sleeping
Upload 6 files
Browse files- DockerFile +19 -0
- env.py +103 -0
- inference.py +84 -0
- models.py +21 -0
- openenv.yaml +16 -0
- requirements.txt +0 -0
DockerFile
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Use a lightweight Python base
|
| 2 |
+
FROM python:3.10-slim
|
| 3 |
+
|
| 4 |
+
# Set working directory
|
| 5 |
+
WORKDIR /app
|
| 6 |
+
|
| 7 |
+
# Copy requirements and install
|
| 8 |
+
COPY requirements.txt .
|
| 9 |
+
RUN pip install --no-cache-dir -r requirements.txt
|
| 10 |
+
|
| 11 |
+
# Copy all project files
|
| 12 |
+
COPY . .
|
| 13 |
+
|
| 14 |
+
# Environment variables for OpenEnv (defaults)
|
| 15 |
+
ENV API_BASE_URL="https://router.huggingface.co/v1"
|
| 16 |
+
ENV MODEL_NAME="Qwen/Qwen2.5-72B-Instruct"
|
| 17 |
+
|
| 18 |
+
# The command to run your inference
|
| 19 |
+
CMD ["python", "inference.py"]
|
env.py
ADDED
|
@@ -0,0 +1,103 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
from typing import Tuple, Dict, Any
|
| 3 |
+
|
| 4 |
+
# Updated imports based on your library's structure
|
| 5 |
+
from openenv.core.env_server import Environment
|
| 6 |
+
from models import Action, Observation, Reward
|
| 7 |
+
|
| 8 |
+
# Change the class name from CodeReviewEnv(OpenEnv) to CodeReviewEnv(Environment)
|
| 9 |
+
class CodeReviewEnv(Environment):
|
| 10 |
+
def __init__(self):
|
| 11 |
+
# In the new version, spec loading is often handled via the CLI
|
| 12 |
+
# or you can just initialize the base class directly.
|
| 13 |
+
super().__init__()
|
| 14 |
+
self.current_task_id = None
|
| 15 |
+
self.code = ""
|
| 16 |
+
self.step_count = 0
|
| 17 |
+
self.max_steps = 5
|
| 18 |
+
|
| 19 |
+
# ... the rest of your methods (reset, step, etc.) remain the same ...
|
| 20 |
+
|
| 21 |
+
def reset(self, task_id: str = "style-cleanup") -> Observation:
|
| 22 |
+
"""Starts a new task session."""
|
| 23 |
+
self.current_task_id = task_id
|
| 24 |
+
self.step_count = 0
|
| 25 |
+
|
| 26 |
+
# Load the buggy code from our data folder
|
| 27 |
+
file_map = {
|
| 28 |
+
"style-cleanup": "data/style.py",
|
| 29 |
+
"efficiency-boost": "data/logic.py",
|
| 30 |
+
"security-audit": "data/security.py"
|
| 31 |
+
}
|
| 32 |
+
|
| 33 |
+
file_path = file_map.get(task_id, "data/style.py")
|
| 34 |
+
with open(file_path, "r") as f:
|
| 35 |
+
self.code = f.read()
|
| 36 |
+
|
| 37 |
+
return self._get_observation()
|
| 38 |
+
|
| 39 |
+
def step(self, action: Action) -> Tuple[Observation, float, bool, Dict[str, Any]]:
|
| 40 |
+
"""Handles the agent's move and returns the reward."""
|
| 41 |
+
self.step_count += 1
|
| 42 |
+
reward_val = 0.0
|
| 43 |
+
done = False
|
| 44 |
+
|
| 45 |
+
# Logic: If the agent provides a fix, update our internal 'code' state
|
| 46 |
+
if action.action_type == "apply_fix":
|
| 47 |
+
self.code = action.content
|
| 48 |
+
reward_val = self._calculate_reward() # Partial progress
|
| 49 |
+
|
| 50 |
+
elif action.action_type == "submit":
|
| 51 |
+
reward_val = self._calculate_reward()
|
| 52 |
+
done = True
|
| 53 |
+
|
| 54 |
+
if self.step_count >= self.max_steps:
|
| 55 |
+
done = True
|
| 56 |
+
|
| 57 |
+
return self._get_observation(), reward_val, done, {}
|
| 58 |
+
|
| 59 |
+
def _get_observation(self) -> Observation:
|
| 60 |
+
"""Returns what the agent sees right now."""
|
| 61 |
+
return Observation(
|
| 62 |
+
file_name=f"{self.current_task_id}.py",
|
| 63 |
+
code_content=self.code,
|
| 64 |
+
diff="Changes applied to buffer.",
|
| 65 |
+
linter_report=self._mock_linter(),
|
| 66 |
+
current_task=self.current_task_id
|
| 67 |
+
)
|
| 68 |
+
|
| 69 |
+
def _mock_linter(self):
|
| 70 |
+
"""A simple linter to guide the agent."""
|
| 71 |
+
errors = []
|
| 72 |
+
if "import sys" in self.code and self.current_task_id == "style-cleanup":
|
| 73 |
+
errors.append("L001: Unused import 'sys' detected.")
|
| 74 |
+
if "f\"SELECT" in self.code and self.current_task_id == "security-audit":
|
| 75 |
+
errors.append("S001: SQL Injection vulnerability detected.")
|
| 76 |
+
return errors
|
| 77 |
+
|
| 78 |
+
def _calculate_reward(self) -> float:
|
| 79 |
+
"""
|
| 80 |
+
CRITICAL: Deterministic grading logic.
|
| 81 |
+
This is where 25% of your score comes from.
|
| 82 |
+
"""
|
| 83 |
+
score = 0.0
|
| 84 |
+
|
| 85 |
+
if self.current_task_id == "style-cleanup":
|
| 86 |
+
# Reward for fixing indentation AND removing the unused import
|
| 87 |
+
if "import sys" not in self.code: score += 0.5
|
| 88 |
+
if " print(\"Indentation" in self.code: score += 0.5
|
| 89 |
+
|
| 90 |
+
elif self.current_task_id == "security-audit":
|
| 91 |
+
# Reward for using parameterized queries instead of f-strings
|
| 92 |
+
if "execute(query)" not in self.code and ("?" in self.code or "," in self.code):
|
| 93 |
+
score = 1.0
|
| 94 |
+
|
| 95 |
+
elif self.current_task_id == "efficiency-boost":
|
| 96 |
+
# Reward for getting rid of the nested loop
|
| 97 |
+
if self.code.count("for ") == 1: # Suggests they used a set or dict
|
| 98 |
+
score = 1.0
|
| 99 |
+
|
| 100 |
+
return min(max(score, 0.0), 1.0) # Ensure it stays in [0, 1]
|
| 101 |
+
|
| 102 |
+
def state(self) -> Dict[str, Any]:
|
| 103 |
+
return {"code": self.code, "task": self.current_task_id}
|
inference.py
ADDED
|
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import json
|
| 3 |
+
import asyncio
|
| 4 |
+
from openai import OpenAI
|
| 5 |
+
from env import CodeReviewEnv
|
| 6 |
+
from models import Action
|
| 7 |
+
|
| 8 |
+
# 1. Setup Environment Variables
|
| 9 |
+
API_KEY = os.getenv("HF_TOKEN")
|
| 10 |
+
API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
|
| 11 |
+
MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
|
| 12 |
+
|
| 13 |
+
client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
|
| 14 |
+
|
| 15 |
+
def log_start(task: str, model: str):
|
| 16 |
+
print(f"[START] task={task} env=aion-code-reviewer model={model}", flush=True)
|
| 17 |
+
|
| 18 |
+
def log_step(step: int, action_type: str, reward: float, done: bool, error: str = "null"):
|
| 19 |
+
done_str = str(done).lower()
|
| 20 |
+
print(f"[STEP] step={step} action={action_type} reward={reward:.2f} done={done_str} error={error}", flush=True)
|
| 21 |
+
|
| 22 |
+
def log_end(success: bool, steps: int, score: float, rewards: list):
|
| 23 |
+
rewards_str = ",".join(f"{r:.2f}" for r in rewards)
|
| 24 |
+
success_str = str(success).lower()
|
| 25 |
+
print(f"[END] success={success_str} steps={steps} score={score:.3f} rewards={rewards_str}", flush=True)
|
| 26 |
+
|
| 27 |
+
async def run_inference():
|
| 28 |
+
env = CodeReviewEnv()
|
| 29 |
+
tasks = ["style-cleanup", "efficiency-boost", "security-audit"]
|
| 30 |
+
|
| 31 |
+
for task_id in tasks:
|
| 32 |
+
obs = env.reset(task_id=task_id)
|
| 33 |
+
log_start(task_id, MODEL_NAME)
|
| 34 |
+
|
| 35 |
+
step_idx = 1
|
| 36 |
+
total_rewards = []
|
| 37 |
+
is_success = False
|
| 38 |
+
|
| 39 |
+
while step_idx <= 5:
|
| 40 |
+
# 2. Build the Prompt for the LLM
|
| 41 |
+
prompt = f"""
|
| 42 |
+
Task: {obs.current_task}
|
| 43 |
+
Current Code:
|
| 44 |
+
{obs.code_content}
|
| 45 |
+
Linter Errors: {obs.linter_report}
|
| 46 |
+
|
| 47 |
+
Return a JSON object with:
|
| 48 |
+
{{
|
| 49 |
+
"action_type": "apply_fix" | "submit",
|
| 50 |
+
"content": "the entire corrected code string",
|
| 51 |
+
"line_number": 0
|
| 52 |
+
}}
|
| 53 |
+
"""
|
| 54 |
+
|
| 55 |
+
try:
|
| 56 |
+
response = client.chat.completions.create(
|
| 57 |
+
model=MODEL_NAME,
|
| 58 |
+
messages=[{"role": "user", "content": prompt}],
|
| 59 |
+
response_format={ "type": "json_object" }
|
| 60 |
+
)
|
| 61 |
+
|
| 62 |
+
# 3. Parse LLM Response
|
| 63 |
+
data = json.loads(response.choices[0].message.content)
|
| 64 |
+
agent_action = Action(**data)
|
| 65 |
+
|
| 66 |
+
# 4. Take Step in Env
|
| 67 |
+
obs, reward, done, info = env.step(agent_action)
|
| 68 |
+
total_rewards.append(reward)
|
| 69 |
+
|
| 70 |
+
log_step(step_idx, agent_action.action_type, reward, done)
|
| 71 |
+
|
| 72 |
+
if done:
|
| 73 |
+
is_success = reward >= 0.8 # Threshold for success
|
| 74 |
+
break
|
| 75 |
+
|
| 76 |
+
step_idx += 1
|
| 77 |
+
except Exception as e:
|
| 78 |
+
log_step(step_idx, "error", 0.0, True, error=str(e))
|
| 79 |
+
break
|
| 80 |
+
|
| 81 |
+
log_end(is_success, step_idx, max(total_rewards) if total_rewards else 0.0, total_rewards)
|
| 82 |
+
|
| 83 |
+
if __name__ == "__main__":
|
| 84 |
+
asyncio.run(run_inference())
|
models.py
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from pydantic import BaseModel, Field
|
| 2 |
+
from typing import Optional, List
|
| 3 |
+
|
| 4 |
+
class Action(BaseModel):
|
| 5 |
+
# The agent chooses what to do
|
| 6 |
+
action_type: str = Field(..., description="Values: 'comment', 'apply_fix', or 'submit'")
|
| 7 |
+
line_number: Optional[int] = Field(None, description="The line number to act upon")
|
| 8 |
+
content: str = Field(..., description="The text of the comment or the code to replace")
|
| 9 |
+
|
| 10 |
+
class Observation(BaseModel):
|
| 11 |
+
# What the agent 'sees' every turn
|
| 12 |
+
file_name: str
|
| 13 |
+
code_content: str
|
| 14 |
+
diff: str
|
| 15 |
+
linter_report: List[str]
|
| 16 |
+
current_task: str
|
| 17 |
+
|
| 18 |
+
class Reward(BaseModel):
|
| 19 |
+
# The score given back to the agent
|
| 20 |
+
value: float = Field(..., ge=0.0, le=1.0)
|
| 21 |
+
comment: str
|
openenv.yaml
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version: "1.0.0"
|
| 2 |
+
name: "aion-code-reviewer"
|
| 3 |
+
description: "An AI-powered sandbox for evaluating code review and bug-fixing capabilities."
|
| 4 |
+
tasks:
|
| 5 |
+
- id: "style-cleanup"
|
| 6 |
+
name: "Style & Linting"
|
| 7 |
+
difficulty: "easy"
|
| 8 |
+
description: "Fix indentation and remove unused imports."
|
| 9 |
+
- id: "efficiency-boost"
|
| 10 |
+
name: "Algorithm Optimization"
|
| 11 |
+
difficulty: "medium"
|
| 12 |
+
description: "Refactor a nested loop into a more efficient lookup."
|
| 13 |
+
- id: "security-audit"
|
| 14 |
+
name: "Security Vulnerability"
|
| 15 |
+
difficulty: "hard"
|
| 16 |
+
description: "Identify and fix a SQL injection vulnerability."
|
requirements.txt
ADDED
|
File without changes
|