File size: 21,331 Bytes
9ccdf0d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 |
#!/usr/bin/env python3
"""
CF-HoT HEAD TRAINING - Contrastive Fine-tuning with Hidden-state Oversight Training
====================================================================================
Trains lightweight "heads" on model hidden states to detect and suppress:
- Repetition (loops, repeated phrases)
- Hedging ("As an AI...", "That's a great question!")
- Verbosity ("Let me explain...", "To put it simply...")
Usage:
python train_cfhot_head.py --behavior repetition --steps 5000
python train_cfhot_head.py --behavior hedging --steps 3000
python train_cfhot_head.py --behavior verbosity --steps 3000
python train_cfhot_head.py --behavior all --steps 3000
"Predict the problem before it happens, prevent it at the source"
"""
import os
import sys
import json
import argparse
import random
from datetime import datetime
from pathlib import Path
from typing import List, Dict, Any, Tuple
from dataclasses import dataclass
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
# === PATHS ===
ROOT = os.path.dirname(os.path.abspath(__file__))
RESULTS_DIR = os.path.join(ROOT, "results")
DATA_DIR = os.path.join(ROOT, "cfhot_data")
os.makedirs(RESULTS_DIR, exist_ok=True)
os.makedirs(DATA_DIR, exist_ok=True)
# Model path - adjust to your setup
MODEL_PATH = "/mnt/nvme2/ubermesnchetien4/models/merged-final-v5"
# ==============================================================================
# DATA GENERATION - POSITIVE AND NEGATIVE EXAMPLES
# ==============================================================================
# REPETITION: Examples that repeat vs don't repeat
REPETITION_POSITIVE = [
# Repeating phrases
"The key is to understand, the key is to understand, the key is to understand that",
"We need to consider, we need to consider, we need to think about",
"It's important to note, it's important to note that this is important to note",
"First, let me say, first let me say, first I want to say",
"The thing is, the thing is, the thing is that we should",
"As I mentioned, as I mentioned before, as I mentioned earlier",
"To be clear, to be clear, to be perfectly clear about this",
"In other words, in other words, to put it another way, in other words",
"The point is, the point is, my point is that the point is",
"What I mean is, what I mean is, what I'm trying to say is what I mean",
# Word repetition
"very very very important",
"really really really good",
"so so so much better",
"the the the problem is",
"I I I think that",
]
REPETITION_NEGATIVE = [
# Clean, varied language
"The key insight here is understanding the underlying mechanism.",
"We should consider multiple perspectives on this issue.",
"This is an important point worth emphasizing.",
"Let me explain the concept clearly.",
"The situation requires careful analysis.",
"First, we examine the data. Then, we draw conclusions.",
"To clarify: the process involves three distinct steps.",
"In simpler terms, the algorithm optimizes for efficiency.",
"The central argument rests on empirical evidence.",
"What this means in practice is significant improvement.",
"Neural networks learn representations automatically.",
"Gradient descent minimizes the loss function iteratively.",
"Recursion solves problems by breaking them into smaller subproblems.",
"Hash tables provide O(1) average-case lookup time.",
"Transformers use attention mechanisms for sequence modeling.",
]
# HEDGING: Sycophantic/apologetic phrases vs direct responses
HEDGING_POSITIVE = [
"That's a great question! Let me think about this.",
"What a fascinating topic! I'd be happy to explore this with you.",
"That's an excellent point! Thank you for bringing this up.",
"I appreciate you asking! This is something I find very interesting.",
"Great question! Many people wonder about this.",
"As an AI language model, I don't have personal experiences, but",
"I apologize, but I'm not able to provide that information.",
"I'm sorry, but I cannot help with that request.",
"Thank you for your patience! Let me try to help.",
"I understand your concern! That's completely valid.",
"What a wonderful question! I'm delighted to assist.",
"I really appreciate you sharing that with me!",
"That's so interesting! Tell me more about that.",
"I'm honored you asked me! Let me do my best.",
"Oh, that's a tricky one! But I'll give it a shot.",
]
HEDGING_NEGATIVE = [
"The answer is straightforward: use a hash table.",
"Recursion works by calling the function with smaller inputs.",
"Neural networks learn through gradient descent.",
"The algorithm has O(n log n) time complexity.",
"This approach fails because it doesn't account for edge cases.",
"The data shows a clear correlation between the variables.",
"Quantum mechanics describes probability amplitudes.",
"Evolution operates through natural selection.",
"The proof follows from the axioms directly.",
"TCP ensures reliable data transmission.",
"Compile the code with optimization flags enabled.",
"The database index improves query performance.",
"Cache invalidation is a hard problem.",
"The gradient points in the direction of steepest ascent.",
"Entropy measures the disorder of a system.",
]
# VERBOSITY: Wordy preambles vs direct starts
VERBOSITY_POSITIVE = [
"Let me explain this to you in detail so you can understand.",
"To put it simply, what I'm trying to say is that",
"In other words, to clarify what I mean, basically",
"First of all, before I answer, I should mention that",
"To begin with, it's important to understand that",
"Essentially, what this boils down to is the fact that",
"Basically, in simple terms, what we're looking at here is",
"Allow me to elaborate on this point for you.",
"I'd like to take a moment to explain this concept.",
"Before we dive in, let me provide some context.",
"To give you a comprehensive answer, I'll need to explain",
"In order to fully understand this, we must first consider",
"The thing you need to know about this is that",
"What you're essentially asking about is related to",
"To answer your question thoroughly, let me start by saying",
]
VERBOSITY_NEGATIVE = [
"Hash tables use O(1) lookup.",
"The gradient points downhill.",
"Recursion needs a base case.",
"Attention weights sum to one.",
"TCP guarantees delivery.",
"Entropy increases over time.",
"Backprop computes gradients.",
"DNA encodes proteins.",
"Light travels at c.",
"Neurons fire or don't.",
"Memory is limited.",
"Caching improves speed.",
"Indexes help queries.",
"Locks prevent races.",
"Tests catch bugs.",
]
# ==============================================================================
# MULTI-HEAD PREDICTOR ARCHITECTURE
# ==============================================================================
class RiskPredictor(nn.Module):
"""Single-head risk predictor for one behavior type."""
def __init__(self, d_model: int, n_layers: int, d_fiber: int = 16, d_control: int = 64):
super().__init__()
self.d_model = d_model
self.n_layers = n_layers
self.d_fiber = d_fiber
# Fiber projections for each layer
self.fiber_projs = nn.ModuleList([
nn.Linear(d_model, d_fiber, bias=False) for _ in range(n_layers)
])
# Learnable layer weights
self.layer_weights = nn.Parameter(torch.ones(n_layers) / n_layers)
# Prediction head
self.predictor = nn.Sequential(
nn.Linear(d_fiber, d_control),
nn.GELU(),
nn.Linear(d_control, d_control),
nn.GELU(),
nn.Linear(d_control, 1)
)
def forward(self, hidden_states: List[torch.Tensor]) -> torch.Tensor:
"""
Args:
hidden_states: List of [batch, seq_len, d_model] tensors, one per layer
Returns:
risk_scores: [batch, seq_len] tensor of risk probabilities
"""
# Project each layer to fiber space
fibers = []
for i, (proj, h) in enumerate(zip(self.fiber_projs, hidden_states)):
if i < len(hidden_states):
fibers.append(proj(h.float()))
# Aggregate with learned weights
weights = F.softmax(self.layer_weights[:len(fibers)], dim=0)
aggregated = sum(w * f for w, f in zip(weights, fibers))
# Predict risk
logits = self.predictor(aggregated).squeeze(-1)
return torch.sigmoid(logits)
class MultiHeadPredictor(nn.Module):
"""Multi-head predictor for all behavior types."""
def __init__(self, d_model: int, n_layers: int, d_fiber: int = 16, d_control: int = 64):
super().__init__()
self.d_model = d_model
self.n_layers = n_layers
self.d_fiber = d_fiber
# Shared fiber projections
self.fiber_projs = nn.ModuleList([
nn.Linear(d_model, d_fiber, bias=False) for _ in range(n_layers)
])
self.layer_weights = nn.Parameter(torch.ones(n_layers) / n_layers)
# Behavior-specific heads
self.heads = nn.ModuleDict({
'repetition': self._make_head(d_fiber, d_control),
'hedging': self._make_head(d_fiber, d_control),
'verbosity': self._make_head(d_fiber, d_control),
})
def _make_head(self, d_fiber: int, d_control: int) -> nn.Module:
return nn.Sequential(
nn.Linear(d_fiber, d_control),
nn.GELU(),
nn.Linear(d_control, d_control),
nn.GELU(),
nn.Linear(d_control, 1)
)
def forward(self, hidden_states: List[torch.Tensor], head_name: str) -> torch.Tensor:
# Project to fiber space
fibers = [proj(h.float()) for proj, h in zip(self.fiber_projs, hidden_states)]
weights = F.softmax(self.layer_weights[:len(fibers)], dim=0)
aggregated = sum(w * f for w, f in zip(weights, fibers))
# Apply specific head
logits = self.heads[head_name](aggregated).squeeze(-1)
return torch.sigmoid(logits)
def get_all_risks(self, hidden_states: List[torch.Tensor]) -> Dict[str, torch.Tensor]:
fibers = [proj(h.float()) for proj, h in zip(self.fiber_projs, hidden_states)]
weights = F.softmax(self.layer_weights[:len(fibers)], dim=0)
aggregated = sum(w * f for w, f in zip(weights, fibers))
return {
name: torch.sigmoid(head(aggregated).squeeze(-1))
for name, head in self.heads.items()
}
# ==============================================================================
# TRAINING
# ==============================================================================
def get_data_for_behavior(behavior: str) -> Tuple[List[str], List[str]]:
"""Get positive and negative examples for a behavior."""
if behavior == "repetition":
return REPETITION_POSITIVE, REPETITION_NEGATIVE
elif behavior == "hedging":
return HEDGING_POSITIVE, HEDGING_NEGATIVE
elif behavior == "verbosity":
return VERBOSITY_POSITIVE, VERBOSITY_NEGATIVE
else:
raise ValueError(f"Unknown behavior: {behavior}")
def collect_hidden_states(model, tokenizer, texts: List[str], device) -> List[torch.Tensor]:
"""Collect hidden states from model for given texts."""
all_hidden_states = []
model.eval()
with torch.no_grad():
for text in texts:
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
inputs = {k: v.to(device) for k, v in inputs.items()}
outputs = model(**inputs, output_hidden_states=True, return_dict=True)
# Get hidden states from all layers [n_layers, batch, seq, d_model]
hidden = outputs.hidden_states[1:] # Skip embedding layer
# Take the last token's hidden state from each layer
last_hidden = [h[:, -1, :] for h in hidden] # [n_layers] of [batch, d_model]
all_hidden_states.append(last_hidden)
return all_hidden_states
def train_head(
behavior: str,
model_path: str,
steps: int = 3000,
lr: float = 1e-4,
d_fiber: int = 16,
d_control: int = 64,
checkpoint_every: int = 500
):
"""Train a single behavior head."""
print(f"\n{'='*70}")
print(f"TRAINING {behavior.upper()} HEAD")
print(f"{'='*70}")
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
# Load model
print(f"[{behavior}] Loading model: {model_path}")
tokenizer = AutoTokenizer.from_pretrained(model_path, local_files_only=True)
tokenizer.pad_token = tokenizer.eos_token
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16,
)
model = AutoModelForCausalLM.from_pretrained(
model_path,
quantization_config=bnb_config,
device_map="auto",
torch_dtype=torch.bfloat16,
local_files_only=True
)
model.eval()
device = next(model.parameters()).device
n_layers = model.config.num_hidden_layers
d_model = model.config.hidden_size
print(f"[{behavior}] Model loaded: {n_layers} layers, {d_model} dims")
# Get training data
positive_texts, negative_texts = get_data_for_behavior(behavior)
print(f"[{behavior}] Data: {len(positive_texts)} positive, {len(negative_texts)} negative")
# Collect hidden states
print(f"[{behavior}] Collecting hidden states...")
positive_hidden = collect_hidden_states(model, tokenizer, positive_texts, device)
negative_hidden = collect_hidden_states(model, tokenizer, negative_texts, device)
# Initialize predictor
predictor = RiskPredictor(d_model, n_layers, d_fiber, d_control).to(device).float()
optimizer = torch.optim.AdamW(predictor.parameters(), lr=lr)
criterion = nn.BCELoss()
# Training loop
predictor.train()
total_loss = 0
results_dir = os.path.join(RESULTS_DIR, f"{behavior}_head")
os.makedirs(results_dir, exist_ok=True)
for step in range(steps):
# Sample batch
if random.random() > 0.5:
# Positive example
idx = random.randint(0, len(positive_hidden) - 1)
hidden = positive_hidden[idx]
target = torch.ones(1, device=device)
else:
# Negative example
idx = random.randint(0, len(negative_hidden) - 1)
hidden = negative_hidden[idx]
target = torch.zeros(1, device=device)
# Forward
pred = predictor(hidden)
pred = pred.mean() # Average over sequence
loss = criterion(pred.unsqueeze(0), target)
# Backward
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(predictor.parameters(), 1.0)
optimizer.step()
total_loss += loss.item()
if (step + 1) % 100 == 0:
avg_loss = total_loss / 100
print(f" Step {step+1}/{steps}: loss={avg_loss:.4f}")
total_loss = 0
# Checkpoint
if (step + 1) % checkpoint_every == 0:
ckpt_dir = os.path.join(results_dir, f"ckpt_{step+1}")
os.makedirs(ckpt_dir, exist_ok=True)
# Evaluate separation
predictor.eval()
with torch.no_grad():
pos_scores = [predictor(h).mean().item() for h in positive_hidden]
neg_scores = [predictor(h).mean().item() for h in negative_hidden]
predictor.train()
avg_pos = sum(pos_scores) / len(pos_scores)
avg_neg = sum(neg_scores) / len(neg_scores)
separation = avg_pos / max(avg_neg, 1e-6)
print(f"\n Checkpoint {step+1}:")
print(f" Avg positive: {avg_pos:.4f}")
print(f" Avg negative: {avg_neg:.4f}")
print(f" Separation: {separation:.1f}x\n")
# Save
torch.save({
'step': step + 1,
'predictor_state': predictor.state_dict(),
'risk_predictor': {
**{f'fiber_projs.{i}.weight': predictor.fiber_projs[i].weight for i in range(n_layers)},
'layer_weights': predictor.layer_weights,
'predictor.0.weight': predictor.predictor[0].weight,
'predictor.0.bias': predictor.predictor[0].bias,
'predictor.2.weight': predictor.predictor[2].weight,
'predictor.2.bias': predictor.predictor[2].bias,
'predictor.4.weight': predictor.predictor[4].weight,
'predictor.4.bias': predictor.predictor[4].bias,
},
'result': {
'avg_positive': avg_pos,
'avg_negative': avg_neg,
'separation': separation,
}
}, os.path.join(ckpt_dir, f"{behavior}_head.pt"))
# Also save as risk_predictor.pt for compatibility
torch.save({
'step': step + 1,
'risk_predictor': {
**{f'fiber_projs.{i}.weight': predictor.fiber_projs[i].weight for i in range(n_layers)},
'layer_weights': predictor.layer_weights,
'predictor.0.weight': predictor.predictor[0].weight,
'predictor.0.bias': predictor.predictor[0].bias,
'predictor.2.weight': predictor.predictor[2].weight,
'predictor.2.bias': predictor.predictor[2].bias,
'predictor.4.weight': predictor.predictor[4].weight,
'predictor.4.bias': predictor.predictor[4].bias,
},
'result': {
'avg_positive': avg_pos,
'avg_negative': avg_neg,
'separation': separation,
}
}, os.path.join(ckpt_dir, "risk_predictor.pt"))
# Final evaluation
predictor.eval()
with torch.no_grad():
pos_scores = [predictor(h).mean().item() for h in positive_hidden]
neg_scores = [predictor(h).mean().item() for h in negative_hidden]
avg_pos = sum(pos_scores) / len(pos_scores)
avg_neg = sum(neg_scores) / len(neg_scores)
separation = avg_pos / max(avg_neg, 1e-6)
print(f"\n{'='*50}")
print(f"FINAL RESULTS - {behavior.upper()} HEAD")
print(f"{'='*50}")
print(f" Avg positive score: {avg_pos:.4f}")
print(f" Avg negative score: {avg_neg:.4f}")
print(f" Separation: {separation:.1f}x")
print(f"{'='*50}")
return {
'behavior': behavior,
'separation': separation,
'avg_positive': avg_pos,
'avg_negative': avg_neg,
'results_dir': results_dir,
}
def train_all_heads(model_path: str, steps: int = 3000):
"""Train all behavior heads."""
results = {}
for behavior in ["repetition", "hedging", "verbosity"]:
result = train_head(behavior, model_path, steps)
results[behavior] = result
print("\n" + "="*70)
print("ALL HEADS TRAINED")
print("="*70)
for behavior, result in results.items():
print(f" {behavior}: {result['separation']:.1f}x separation")
print("="*70)
return results
# ==============================================================================
# MAIN
# ==============================================================================
def main():
parser = argparse.ArgumentParser(description="CF-HoT Head Training")
parser.add_argument("--behavior", type=str, default="repetition",
help="Behavior to train: repetition, hedging, verbosity, all")
parser.add_argument("--steps", type=int, default=3000, help="Training steps")
parser.add_argument("--lr", type=float, default=1e-4, help="Learning rate")
parser.add_argument("--model-path", type=str, default=MODEL_PATH, help="Base model path")
parser.add_argument("--d-fiber", type=int, default=16, help="Fiber dimension")
parser.add_argument("--d-control", type=int, default=64, help="Control dimension")
args = parser.parse_args()
print("="*70)
print("CF-HoT HEAD TRAINING")
print("="*70)
print(f" Behavior: {args.behavior}")
print(f" Steps: {args.steps}")
print(f" Learning rate: {args.lr}")
print(f" Model: {args.model_path}")
print("="*70)
if args.behavior == "all":
train_all_heads(args.model_path, args.steps)
else:
train_head(
args.behavior,
args.model_path,
args.steps,
args.lr,
args.d_fiber,
args.d_control
)
if __name__ == "__main__":
main()
|