File size: 11,004 Bytes
60ddd88 | 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 | """
Train H4 cross-encoder reranker on SQuAD.
Uses the PPL 10.0 TinyStories checkpoint as backbone.
Fine-tunes on binary classification: does this passage answer this question?
For each SQuAD example:
- Positive: [question SEP correct_passage] -> label 1
- Negative: [question SEP wrong_passage] -> label 0
The H4 attention heads directly attend from question tokens to passage tokens
within the same sequence — this is why cross-encoders beat bi-encoders.
Pipeline integration:
1. Bi-encoder retrieves top-5 (R@5=100%, 20ms)
2. Cross-encoder reranks 5 candidates (5 forward passes, ~50ms)
3. Return highest-scoring → R@1 should reach 80-90%+
"""
import os
import math
import time
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import sys
sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
from rag.cross_encoder import H4CrossEncoder
from rag.prepare_qa import download_squad_dev
from rag.tokenizer import BPETokenizer
# ---------------------------------------------------------------------------
# Hyperparameters
# ---------------------------------------------------------------------------
TIME_BUDGET = int(os.environ.get('CE_TIME', 3600)) # 1 hour default
D_MODEL = 512
N_HEADS = 8
N_LAYERS = 8
USE_BITLINEAR = True
LR = 5e-4 # lower LR for fine-tuning (backbone is pre-trained)
WEIGHT_DECAY = 0.01
GRAD_CLIP = 1.0
BATCH_SIZE = 8 # pairs per batch (each has 1 positive + 1 negative)
MAX_SEQ_LEN = 192 # question + passage combined
EVAL_INTERVAL = 100
CHECKPOINT_PATH = os.path.join(os.path.dirname(__file__), '..', '..', 'checkpoints', 'h4_fullscale_final.pt')
def pad_tokens(ids, max_len):
ids = ids[:max_len]
return ids + [0] * (max_len - len(ids))
def make_pair(tokenizer, question, passage, max_len):
"""Encode [question SEP passage] as a single sequence."""
q_ids = tokenizer.encode(question)
p_ids = tokenizer.encode(passage)
# Budget: half for question, half for passage (with SEP)
max_q = max_len // 3
max_p = max_len - max_q - 1
q_ids = q_ids[:max_q]
p_ids = p_ids[:max_p]
combined = q_ids + [2] + p_ids # 2 = SEP
return pad_tokens(combined, max_len)
def main():
t_start = time.time()
torch.manual_seed(42)
random.seed(42)
np.random.seed(42)
# Load SQuAD
squad = download_squad_dev()
if len(squad) < 100:
print("SQuAD not available. Run: python rag/prepare_qa.py")
return
print(f"SQuAD: {len(squad)} QA pairs")
# Build BPE tokenizer
tokenizer = BPETokenizer(max_vocab=8192)
all_texts = [qa['context'] + ' ' + qa['question'] for qa in squad[:2000]]
tokenizer.build_vocab(all_texts)
# Split
indices = list(range(len(squad)))
random.shuffle(indices)
n_val = 200
train_data = [squad[i] for i in indices[n_val:]]
val_data = [squad[i] for i in indices[:n_val]]
print(f"Train: {len(train_data)}, Val: {len(val_data)}")
# Create cross-encoder
model = H4CrossEncoder(
vocab_size=tokenizer.vocab_size,
d_model=D_MODEL,
n_heads=N_HEADS,
n_layers=N_LAYERS,
use_bitlinear=USE_BITLINEAR,
max_seq_len=MAX_SEQ_LEN,
)
print(f"Model: {model.count_params():,} params")
# Load pre-trained backbone
if os.path.exists(CHECKPOINT_PATH):
config = model.load_lm_backbone(CHECKPOINT_PATH)
print(f"Loaded backbone from {CHECKPOINT_PATH}")
else:
print(f"No checkpoint at {CHECKPOINT_PATH}, training from scratch")
# Freeze backbone initially, only train score head
# Then unfreeze after warmup for fine-tuning
for name, param in model.lm.named_parameters():
param.requires_grad = False
trainable_head = sum(p.numel() for p in model.score_head.parameters())
print(f"Phase 1: training score head only ({trainable_head:,} params)")
optimizer = torch.optim.AdamW(
filter(lambda p: p.requires_grad, model.parameters()),
lr=LR * 10, # higher LR for head-only phase
weight_decay=WEIGHT_DECAY,
)
# Training loop
model.train()
step = 0
total_training_time = 0.0
best_acc = 0.0
unfrozen = False
UNFREEZE_STEP = 200
print(f"\nTraining for {TIME_BUDGET}s")
print(f"{'step':>6} {'loss':>8} {'acc':>8} {'val_acc':>8} {'phase':>10}")
print("-" * 48)
while True:
t0 = time.time()
# Unfreeze backbone after warmup
if step == UNFREEZE_STEP and not unfrozen:
for param in model.lm.parameters():
param.requires_grad = True
unfrozen = True
total_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f"\n Phase 2: unfreezing backbone ({total_trainable:,} trainable params)")
optimizer = torch.optim.AdamW(
model.parameters(), lr=LR, weight_decay=WEIGHT_DECAY, betas=(0.9, 0.95))
# Sample batch: positive and negative pairs
batch_qa = random.sample(train_data, min(BATCH_SIZE, len(train_data)))
input_ids = []
labels = []
for qa in batch_qa:
# Positive: question + correct passage
pos = make_pair(tokenizer, qa['question'], qa['context'], MAX_SEQ_LEN)
input_ids.append(pos)
labels.append(1.0)
# Negative: question + random wrong passage
neg_qa = random.choice(train_data)
while neg_qa['context'] == qa['context']:
neg_qa = random.choice(train_data)
neg = make_pair(tokenizer, qa['question'], neg_qa['context'], MAX_SEQ_LEN)
input_ids.append(neg)
labels.append(0.0)
input_ids = torch.tensor(input_ids, dtype=torch.long)
labels = torch.tensor(labels, dtype=torch.float32)
# Forward
scores = model(input_ids)
loss = F.binary_cross_entropy_with_logits(scores, labels)
# Accuracy
with torch.no_grad():
preds = (scores > 0).float()
acc = (preds == labels).float().mean().item()
optimizer.zero_grad()
loss.backward()
if GRAD_CLIP > 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), GRAD_CLIP)
optimizer.step()
dt = time.time() - t0
if step > 2:
total_training_time += dt
# Eval
if step % EVAL_INTERVAL == 0:
model.eval()
val_correct = 0
val_total = 0
val_r1 = 0
val_r1_total = 0
with torch.no_grad():
# Binary accuracy
for vi in range(0, min(len(val_data), 100), BATCH_SIZE):
vbatch = val_data[vi:vi + BATCH_SIZE]
v_ids = []
v_labels = []
for qa in vbatch:
pos = make_pair(tokenizer, qa['question'], qa['context'], MAX_SEQ_LEN)
v_ids.append(pos)
v_labels.append(1.0)
neg_qa = random.choice(val_data)
neg = make_pair(tokenizer, qa['question'], neg_qa['context'], MAX_SEQ_LEN)
v_ids.append(neg)
v_labels.append(0.0)
v_ids = torch.tensor(v_ids, dtype=torch.long)
v_labels = torch.tensor(v_labels)
v_scores = model(v_ids)
v_preds = (v_scores > 0).float()
val_correct += (v_preds == v_labels).sum().item()
val_total += len(v_labels)
# Reranking accuracy (R@1 on top-5 candidates)
for qa in val_data[:50]:
# Simulate: 1 correct + 4 wrong passages
candidates = [qa['context']]
neg_pool = [q for q in val_data if q['context'] != qa['context']]
for neg in random.sample(neg_pool, min(4, len(neg_pool))):
candidates.append(neg['context'])
c_ids = []
for passage in candidates:
c_ids.append(make_pair(tokenizer, qa['question'], passage, MAX_SEQ_LEN))
c_ids = torch.tensor(c_ids, dtype=torch.long)
c_scores = model(c_ids)
top_idx = c_scores.argmax().item()
if top_idx == 0: # correct passage was ranked first
val_r1 += 1
val_r1_total += 1
val_acc = val_correct / max(val_total, 1)
rerank_r1 = val_r1 / max(val_r1_total, 1)
if val_acc > best_acc:
best_acc = val_acc
phase = "head-only" if not unfrozen else "full"
print(f"{step:6d} {loss.item():8.4f} {acc:8.3f} {val_acc:8.3f} {phase:>10}"
f" rerank_R@1={rerank_r1:.3f}")
model.train()
step += 1
elapsed = time.time() - t_start
if step > 2 and total_training_time >= TIME_BUDGET:
break
# Final evaluation
model.eval()
print("\n" + "=" * 60)
print("FINAL CROSS-ENCODER EVALUATION:")
final_r1 = 0
final_total = 0
with torch.no_grad():
for qa in val_data[:100]:
candidates = [qa['context']]
neg_pool = [q for q in val_data if q['context'] != qa['context']]
for neg in random.sample(neg_pool, min(4, len(neg_pool))):
candidates.append(neg['context'])
c_ids = []
for passage in candidates:
c_ids.append(make_pair(tokenizer, qa['question'], passage, MAX_SEQ_LEN))
c_ids = torch.tensor(c_ids, dtype=torch.long)
c_scores = model(c_ids)
if c_scores.argmax().item() == 0:
final_r1 += 1
final_total += 1
rerank_r1 = final_r1 / max(final_total, 1)
print(f" Rerank R@1 (top-5): {rerank_r1:.1%} ({final_r1}/{final_total})")
print(f" Best binary acc: {best_acc:.1%}")
print("=" * 60)
print("\n---")
print(f"rerank_r1: {rerank_r1:.4f}")
print(f"best_binary_acc: {best_acc:.4f}")
print(f"training_seconds: {total_training_time:.1f}")
print(f"total_seconds: {time.time() - t_start:.1f}")
print(f"num_steps: {step}")
print(f"num_params: {model.count_params()}")
print(f"ternary: {'yes' if USE_BITLINEAR else 'no'}")
# Save checkpoint
os.makedirs('checkpoints', exist_ok=True)
ckpt_path = os.path.join('checkpoints', 'h4_cross_encoder.pt')
torch.save({
'model_state': model.state_dict(),
'rerank_r1': rerank_r1,
'step': step,
'config': {
'd_model': D_MODEL, 'n_layers': N_LAYERS, 'n_heads': N_HEADS,
'vocab_size': tokenizer.vocab_size, 'use_bitlinear': USE_BITLINEAR,
},
}, ckpt_path)
print(f"Saved: {ckpt_path}")
if __name__ == '__main__':
main()
|