File size: 10,544 Bytes
73400c8 | 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 | #!/usr/bin/env python3
"""
GRPO Training - The Reasoning Magic
Uses the trained model from stage 1
"""
import sys
import json
import torch
import torch.nn as nn
import torch.nn.functional as F
from pathlib import Path
from tqdm import tqdm
sys.path.insert(0, str(Path(__file__).parent.parent))
from src.shorekeeper import SHOREKEEPER, MemoryEfficientSHOREKEEPER
from transformers import AutoTokenizer
class GRPOTrainer:
"""Group Relative Policy Optimization Trainer"""
def __init__(self, model, tokenizer, config):
self.model = model
self.tokenizer = tokenizer
self.device = next(model.parameters()).device
self.group_size = config.get('group_size', 2)
self.lr = config.get('learning_rate', 1e-6)
self.optimizer = torch.optim.AdamW(
self.model.parameters(),
lr=self.lr,
weight_decay=0.01
)
self.step = 0
def compute_reward(self, response, ground_truth):
"""Calculate reward for a response"""
reward = 0.0
# Format reward - check for reasoning tokens
if '|special_token|' in response:
reward += 0.5
# Extract answer (look for numbers at the end)
import re
numbers = re.findall(r'\d+', response)
if numbers:
last_num = numbers[-1]
if last_num == str(ground_truth).strip():
reward += 2.0
# Length reward - not too short
if len(response.split()) > 10:
reward += 0.2
# No repetition penalty
words = response.split()
unique_ratio = len(set(words)) / max(len(words), 1)
if unique_ratio > 0.5:
reward += 0.3
return reward
def generate_response(self, prompt, max_length=128):
"""Generate a response from the model"""
self.model.eval()
try:
inputs = self.tokenizer(prompt, return_tensors="pt", truncation=True, max_length=256)
inputs = {k: v.to(self.device) for k, v in inputs.items()}
with torch.no_grad():
outputs = self.model.generate(
inputs['input_ids'],
max_new_tokens=max_length,
temperature=0.8,
do_sample=True,
pad_token_id=self.tokenizer.eos_token_id
)
response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
return response
except Exception as e:
return f"Error: {e}"
def train_step(self, prompt, ground_truth):
"""Single GRPO step"""
self.model.train()
# Generate group of responses
responses = []
rewards = []
for _ in range(self.group_size):
response = self.generate_response(prompt)
responses.append(response)
reward = self.compute_reward(response, ground_truth)
rewards.append(reward)
# Calculate advantages (relative to group mean)
mean_reward = sum(rewards) / len(rewards)
advantages = [r - mean_reward for r in rewards]
# Train on responses with positive advantage
total_loss = 0
valid_steps = 0
for i, (response, advantage) in enumerate(zip(responses, advantages)):
if advantage <= 0:
continue
# Create training text
text = f"{prompt}\n{response}"
inputs = self.tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
inputs = {k: v.to(self.device) for k, v in inputs.items()}
# Forward pass
logits = self.model(inputs['input_ids'])
# Calculate language modeling loss
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = inputs['input_ids'][..., 1:].contiguous()
loss = F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
ignore_index=self.tokenizer.pad_token_id
)
# Weight by advantage
total_loss = total_loss + loss * advantage
valid_steps += 1
if valid_steps > 0 and total_loss != 0:
total_loss = total_loss / valid_steps
self.optimizer.zero_grad()
total_loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.optimizer.step()
return {
'loss': total_loss.item(),
'avg_reward': sum(rewards) / len(rewards),
'best_reward': max(rewards),
'valid_steps': valid_steps
}
return {
'loss': 0,
'avg_reward': sum(rewards) / len(rewards),
'best_reward': max(rewards),
'valid_steps': 0
}
def train(self, dataset, num_epochs=1):
"""Full training loop"""
print(f"\nTraining on device: {self.device}")
for epoch in range(num_epochs):
print(f"\n{'='*50}")
print(f"Epoch {epoch + 1}/{num_epochs}")
print(f"{'='*50}")
total_loss = 0
total_reward = 0
steps = 0
valid_steps = 0
pbar = tqdm(dataset, desc=f"GRPO Training")
for i, item in enumerate(pbar):
prompt = item.get('prompt', '')
answer = item.get('answer', item.get('ground_truth', ''))
if not prompt or not answer:
continue
try:
stats = self.train_step(prompt, str(answer))
if stats['valid_steps'] > 0:
total_loss += stats['loss']
valid_steps += 1
total_reward += stats['avg_reward']
steps += 1
pbar.set_postfix({
'loss': f'{stats["loss"]:.4f}',
'reward': f'{stats["avg_reward"]:.2f}'
})
except Exception as e:
if i < 10:
print(f"\n Error: {e}")
continue
if steps > 0:
avg_loss = total_loss / valid_steps if valid_steps > 0 else 0
avg_reward = total_reward / steps
print(f"\n Epoch complete: Avg Loss={avg_loss:.4f}, Avg Reward={avg_reward:.2f}")
return self.model
def load_training_data(data_path, limit=None):
"""Load training data for GRPO"""
data = []
data_path = Path(data_path)
if not data_path.exists():
print(f"Data file not found: {data_path}")
return data
with open(data_path, 'r') as f:
for i, line in enumerate(f):
if limit and i >= limit:
break
try:
item = json.loads(line)
data.append({
'prompt': item.get('prompt', ''),
'answer': item.get('ground_truth', item.get('response', ''))
})
except:
continue
return data
def main():
print("=" * 60)
print("SHOREKEEPER GRPO Training")
print("The Reasoning Magic")
print("=" * 60)
# Check device
if torch.cuda.is_available():
device = torch.device("cuda")
print(f"\n✓ CUDA: {torch.cuda.get_device_name(0)}")
print(f" Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")
# Load trained model (full precision for training)
print("\n1. Loading trained SHOREKEEPER model...")
model_path = Path("./outputs/shorekeeper-4b-final.pt")
if not model_path.exists():
print(f"\n❌ Model not found at {model_path}")
print(" Run training first: python3 scripts/04_train.py")
return
model = SHOREKEEPER() # Use full model (not memory efficient for training)
model.load_state_dict(torch.load(model_path, map_location=device))
model = model.to(device)
model.train()
print(f" ✓ Model loaded from {model_path}")
# Load tokenizer
print("\n2. Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained("gpt2")
tokenizer.pad_token = tokenizer.eos_token
print(" ✓ Using GPT-2 tokenizer")
# Load training data
print("\n3. Loading training data...")
data_path = Path("./data/processed/train.jsonl")
if not data_path.exists():
print(f"\n❌ No data at {data_path}")
return
print(" Options:")
print(" [1] Quick test (20 examples)")
print(" [2] Small training (100 examples, 3 epochs)")
choice = input("\nChoose option (1/2): ").strip()
if choice == "1":
limit = 20
epochs = 1
else:
limit = 100
epochs = 3
data = load_training_data(data_path, limit=limit)
print(f"\n Loaded {len(data)} examples")
print(f" Training for {epochs} epochs")
# GRPO config
config = {
'group_size': 2,
'learning_rate': 1e-6
}
print("\n4. Initializing GRPO Trainer...")
trainer = GRPOTrainer(model, tokenizer, config)
print("\n5. Starting GRPO training...")
print(" (This teaches the model to reason)\n")
try:
trained_model = trainer.train(data, num_epochs=epochs)
except KeyboardInterrupt:
print("\n Interrupted")
except Exception as e:
print(f"\n Error: {e}")
import traceback
traceback.print_exc()
# Save model
print("\n6. Saving model...")
output_dir = Path("./outputs/grpo")
output_dir.mkdir(parents=True, exist_ok=True)
torch.save(model.state_dict(), output_dir / "shorekeeper-4b-grpo.pt")
print(f" ✓ Saved to {output_dir / 'shorekeeper-4b-grpo.pt'}")
print("\n" + "=" * 60)
print("✅ GRPO Complete!")
print("=" * 60)
print("\nNow run SHOREKEEPER:")
print(" python3 scripts/07_run_shorekeeper.py")
if __name__ == "__main__":
main()
|