Spaces:
Sleeping
Sleeping
File size: 18,865 Bytes
03a7eb9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 | #!/usr/bin/env python3
"""
Full RL Training Loop for CodeArena with Memory and Fine-tuning
Implements experience replay, trajectory learning, and optimization.
"""
import os
import json
import time
import random
import requests
from typing import List, Dict, Tuple, Optional
from collections import deque
import numpy as np
from dataclasses import dataclass
from datetime import datetime
@dataclass
class Experience:
"""RL Experience tuple"""
state: str # Buggy code + error log + test results
action: str # Generated fix
reward: float
next_state: str
done: bool
task_id: str
step_count: int
trajectory_id: str
@dataclass
class Trajectory:
"""Complete episode trajectory"""
trajectory_id: str
task_id: str
steps: List[Experience]
final_reward: float
success: bool
total_steps: int
class CodeArenaRLTrainer:
def __init__(self, model_name: str = "llama3.2:latest", memory_size: int = 1000):
self.model_name = model_name
self.api_base = "http://localhost:11434"
# RL Components
self.memory = deque(maxlen=memory_size)
self.trajectories: List[Trajectory] = []
self.successful_trajectories: List[Trajectory] = []
# Training parameters
self.learning_rate = 0.001
self.gamma = 0.95 # Discount factor
self.epsilon = 1.0 # Exploration rate
self.epsilon_min = 0.1
self.epsilon_decay = 0.995
self.batch_size = 32
# Task progression
self.current_difficulty = "easy"
self.task_performance = {"easy": [], "medium": [], "hard": []}
# Optimization
self.cache = {} # Response cache for speed
self.prompt_templates = self._load_prompt_templates()
def _load_prompt_templates(self) -> Dict[str, str]:
"""Load optimized prompt templates"""
return {
"base": """You are an expert Python debugger. Fix the buggy code below.
BUGGY CODE:
{buggy_code}
CURRENT ERRORS:
{error_log}
TEST RESULTS:
{test_results}
REQUIREMENTS:
1. The code must compile without syntax errors
2. All tests must pass
3. Fix the ROOT CAUSE, not just symptoms
4. Do NOT repeat previous failed approaches
5. Ensure proper Python syntax and indentation
6. Return ONLY the corrected code, no explanations
Output the complete corrected Python code:""",
"rl_enhanced": """You are learning to debug Python code through reinforcement learning.
PREVIOUS EXPERIENCES:
{similar_experiences}
BUGGY CODE:
{buggy_code}
CURRENT ERRORS:
{error_log}
TEST RESULTS:
{test_results}
LEARNING OBJECTIVE:
- Learn from successful patterns in similar problems
- Avoid mistakes that led to low rewards
- Build upon working solutions
Output ONLY the corrected Python code:""",
"step_aware": """Step {step_count} of debugging process.
{context}
BUGGY CODE:
{buggy_code}
CURRENT ERRORS:
{error_log}
TEST RESULTS:
{test_results}
STEP {step_count} FOCUS:
{step_instruction}
Output ONLY the corrected Python code:"""
}
def get_similar_experiences(self, current_state: str, limit: int = 3) -> str:
"""Retrieve similar successful experiences from memory"""
if not self.successful_trajectories:
return "No previous successful experiences available."
# Simple similarity based on code length and error patterns
current_length = len(current_state)
similar = []
for traj in self.successful_trajectories[-10:]: # Last 10 successful
for exp in traj.steps:
if exp.reward > 0.5: # Only successful steps
length_diff = abs(len(exp.state) - current_length)
if length_diff < 200: # Similar complexity
similar.append(f"β Success: {exp.action[:100]}... (reward: {exp.reward:.2f})")
if len(similar) >= limit:
break
if len(similar) >= limit:
break
return "\n".join(similar) if similar else "Learning from general patterns..."
def generate_fix_rl(self, buggy_code: str, error_log: str, test_results: str,
previous_attempts: List[str], step_count: int,
use_memory: bool = True) -> str:
"""Generate fix using RL-enhanced prompting"""
# Build state representation
state = f"Code: {buggy_code}\nErrors: {error_log}\nTests: {test_results}"
# Choose prompt strategy based on experience
if use_memory and len(self.successful_trajectories) > 0:
similar_exp = self.get_similar_experiences(state)
prompt = self.prompt_templates["rl_enhanced"].format(
similar_experiences=similar_exp,
buggy_code=buggy_code,
error_log=error_log,
test_results=test_results
)
else:
# Step-aware prompting
step_instructions = {
1: "Focus on fixing syntax errors and basic compilation issues first.",
2: "Address logic errors from the previous attempt.",
3: "Optimize and ensure all edge cases are handled.",
4: "Final verification - ensure robust solution.",
5: "Last attempt - use completely different approach if needed."
}
context = ""
if previous_attempts:
context = f"Previous failed attempts:\n" + "\n".join(
f"- {attempt[:50]}..." for attempt in previous_attempts[-2:]
)
prompt = self.prompt_templates["step_aware"].format(
step_count=step_count,
context=context,
buggy_code=buggy_code,
error_log=error_log,
test_results=test_results,
step_instruction=step_instructions.get(step_count, "Fix all issues.")
)
# Check cache first
cache_key = hash(prompt)
if cache_key in self.cache:
return self.cache[cache_key]
try:
response = requests.post(
f"{self.api_base}/api/generate",
json={
"model": self.model_name,
"prompt": prompt,
"stream": False,
"options": {
"temperature": max(0.1, self.epsilon), # Exploration vs exploitation
"num_predict": 800,
"top_p": 0.9
}
},
timeout=20
)
response.raise_for_status()
result = response.json()
fix = result.get("response", "").strip()
# Clean up response
if fix.startswith("```python"):
fix = fix[9:]
if fix.startswith("```"):
fix = fix[3:]
if fix.endswith("```"):
fix = fix[:-3]
fix = fix.strip()
# Cache successful responses
if fix and len(fix) > 10:
self.cache[cache_key] = fix
return fix
except Exception as e:
print(f"API Error: {e}")
return self._fallback_fix(buggy_code, step_count)
def _fallback_fix(self, buggy_code: str, step_count: int) -> str:
"""Enhanced fallback with learning from memory"""
# Try to learn from successful patterns
if self.successful_trajectories:
# Use patterns from successful trajectories
successful_fixes = []
for traj in self.successful_trajectories[-3:]:
for exp in traj.steps:
if exp.reward > 0.6:
successful_fixes.append(exp.action)
if successful_fixes:
# Return a variation of successful fix
base_fix = random.choice(successful_fixes)
# Simple variation - could be improved
return base_fix
# Basic fallback
return "def placeholder_function(x):\n return x"
def run_episode_rl(self, task_id: str, max_steps: int = 5,
use_memory: bool = True) -> Trajectory:
"""Run a single RL episode with memory"""
trajectory_id = f"{task_id}_{int(time.time())}"
print(f"\nπ― RL Episode: {task_id} (Ξ΅={self.epsilon:.3f})")
# Reset environment
try:
response = requests.post("http://localhost:7860/reset",
json={"task_id": task_id}, timeout=10)
response.raise_for_status()
obs = response.json()
except Exception as e:
print(f"β Reset failed: {e}")
return Trajectory(trajectory_id, task_id, [], 0.0, False, 0)
experiences = []
previous_attempts = []
done = False
step_count = 0
final_reward = 0.0
while not done and step_count < max_steps:
step_count += 1
# Build current state
current_state = f"{obs.get('buggy_code', '')}|{obs.get('error_log', '')}|{obs.get('test_results', '')}"
# Generate action using RL
fix = self.generate_fix_rl(
buggy_code=obs.get('buggy_code', ''),
error_log=obs.get('error_log', ''),
test_results=obs.get('test_results', ''),
previous_attempts=previous_attempts,
step_count=step_count,
use_memory=use_memory
)
print(f"π§ Step {step_count}: Generated fix ({len(fix)} chars)")
# Execute action
try:
response = requests.post("http://localhost:7860/step",
json={"proposed_fix": fix}, timeout=20)
response.raise_for_status()
result = response.json()
reward = result.get('reward', 0)
done = result.get('done', False)
next_obs = result.get('observation', {})
# Build next state
next_state = f"{next_obs.get('buggy_code', '')}|{next_obs.get('error_log', '')}|{next_obs.get('test_results', '')}"
# Create experience
exp = Experience(
state=current_state,
action=fix,
reward=reward,
next_state=next_state,
done=done,
task_id=task_id,
step_count=step_count,
trajectory_id=trajectory_id
)
experiences.append(exp)
self.memory.append(exp)
previous_attempts.append(fix)
final_reward = reward
info = result.get('info', {})
print(f" Reward: {reward:.3f}")
print(f" Tests: {info.get('test_results', 'unknown')}")
print(f" Done: {done}")
if reward > 0.5:
print("π Good reward! Learning...")
elif reward < 0.1:
print("β οΈ Low reward - adjusting strategy")
obs = next_obs
except Exception as e:
print(f"β Step failed: {e}")
break
# Create trajectory
success = final_reward > 0.5
trajectory = Trajectory(
trajectory_id=trajectory_id,
task_id=task_id,
steps=experiences,
final_reward=final_reward,
success=success,
total_steps=step_count
)
self.trajectories.append(trajectory)
if success:
self.successful_trajectories.append(trajectory)
# Update task performance
difficulty = task_id.split('-')[0]
if difficulty in self.task_performance:
self.task_performance[difficulty].append(final_reward)
# Decay exploration
self.epsilon = max(self.epsilon_min, self.epsilon * self.epsilon_decay)
print(f"π Episode complete: {success} (reward: {final_reward:.3f})")
return trajectory
def should_progress_difficulty(self) -> Optional[str]:
"""Check if agent should move to next difficulty level"""
if self.current_difficulty == "easy":
recent_easy = self.task_performance["easy"][-3:] # Last 3 episodes
if len(recent_easy) >= 3 and np.mean(recent_easy) > 0.75:
return "medium"
elif self.current_difficulty == "medium":
recent_medium = self.task_performance["medium"][-3:]
if len(recent_medium) >= 3 and np.mean(recent_medium) > 0.70:
return "hard"
return None
def train_rl(self, episodes: int = 50, checkpoint_every: int = 10):
"""Full RL training loop"""
print("π Starting RL Training")
print("=" * 60)
print(f"Model: {self.model_name}")
print(f"Episodes: {episodes}")
print(f"Memory size: {len(self.memory)}")
print(f"Successful trajectories: {len(self.successful_trajectories)}")
results = []
for episode in range(episodes):
# Adaptive task selection
next_difficulty = self.should_progress_difficulty()
if next_difficulty:
self.current_difficulty = next_difficulty
print(f"π Progressing to {self.current_difficulty} difficulty!")
# Select task based on current difficulty
task_candidates = [f"{self.current_difficulty}-{i}" for i in range(1, 4)]
task_id = random.choice(task_candidates)
# Run episode
trajectory = self.run_episode_rl(task_id, use_memory=True)
results.append({
"episode": episode + 1,
"task_id": trajectory.task_id,
"reward": trajectory.final_reward,
"success": trajectory.success,
"steps": trajectory.total_steps,
"epsilon": self.epsilon
})
# Checkpoint
if (episode + 1) % checkpoint_every == 0:
self.save_checkpoint(f"checkpoint_{episode + 1}.json")
print(f"πΎ Checkpoint saved at episode {episode + 1}")
# Performance summary
if (episode + 1) % 10 == 0:
recent_results = results[-10:]
success_rate = sum(1 for r in recent_results if r["success"]) / len(recent_results)
avg_reward = sum(r["reward"] for r in recent_results) / len(recent_results)
print(f"π Episode {episode + 1:3d} | Success: {success_rate:.1%} | Reward: {avg_reward:.3f}")
# Final summary
self.print_training_summary(results)
return results
def print_training_summary(self, results: List[Dict]):
"""Print comprehensive training summary"""
print("\n" + "=" * 60)
print("π RL TRAINING COMPLETE")
print("=" * 60)
total_episodes = len(results)
successful_episodes = sum(1 for r in results if r["success"])
success_rate = successful_episodes / total_episodes
rewards = [r["reward"] for r in results]
avg_reward = np.mean(rewards)
max_reward = max(rewards)
print(f"π Overall Performance:")
print(f"π― Episodes: {total_episodes}")
print(f"β
Successful: {successful_episodes}")
print(f"π Success Rate: {success_rate:.1%}")
print(f"π° Average Reward: {avg_reward:.3f}")
print(f"π Max Reward: {max_reward:.3f}")
print(f"π― Success Rate: {success_rate:.1%}")
# Performance by difficulty
print(f"\nπ Performance by Difficulty:")
for difficulty in ["easy", "medium", "hard"]:
diff_results = [r for r in results if r["task_id"].startswith(difficulty)]
if diff_results:
diff_success = sum(1 for r in diff_results if r["success"]) / len(diff_results)
diff_avg_reward = np.mean([r["reward"] for r in diff_results])
print(f" {difficulty.capitalize()}: Success {diff_success:.1%} | Reward {diff_avg_reward:.3f}")
# Learning curve
print(f"\nπ Learning Curve (last 20 episodes):")
recent = results[-20:]
if recent:
for i in range(0, len(recent), 5):
batch = recent[i:i+5]
batch_success = sum(1 for r in batch if r["success"]) / len(batch)
batch_avg_reward = np.mean([r["reward"] for r in batch])
print(f" Ep {i+1:2d}-{min(i+5, len(recent)):2d}: Success {batch_success:.1%} | Reward {batch_avg_reward:.3f}")
print(f"\nπΎ Memory: {len(self.memory)} experiences")
print(f"ποΈ Successful trajectories: {len(self.successful_trajectories)}")
print(f"π§ Cache size: {len(self.cache)} responses")
def save_checkpoint(self, filename: str):
"""Save training checkpoint"""
checkpoint = {
"timestamp": datetime.now().isoformat(),
"model_name": self.model_name,
"memory_size": len(self.memory),
"successful_trajectories": len(self.successful_trajectories),
"current_difficulty": self.current_difficulty,
"epsilon": self.epsilon,
"task_performance": self.task_performance,
"cache_size": len(self.cache)
}
with open(filename, 'w') as f:
json.dump(checkpoint, f, indent=2)
def main():
import argparse
parser = argparse.ArgumentParser(description="Full RL Training for CodeArena")
parser.add_argument("--episodes", type=int, default=30, help="Number of training episodes")
parser.add_argument("--model", default="llama3.2:latest", help="Ollama model to use")
parser.add_argument("--memory", type=int, default=500, help="Experience replay memory size")
parser.add_argument("--checkpoint", type=int, default=10, help="Save checkpoint every N episodes")
args = parser.parse_args()
print("π§ CodeArena RL Trainer")
print("=" * 50)
print(f"Model: {args.model}")
print(f"Episodes: {args.episodes}")
print(f"Memory: {args.memory}")
print(f"Checkpoints: every {args.checkpoint} episodes")
trainer = CodeArenaRLTrainer(args.model, args.memory)
results = trainer.train_rl(args.episodes, args.checkpoint)
# Save final results
with open("rl_training_results.json", 'w') as f:
json.dump(results, f, indent=2)
print("")
print("πΎ Results saved to rl_training_results.json")
print("π Run 'python plot_rewards.py' to visualize performance")
if __name__ == "__main__":
main() |