File size: 15,378 Bytes
bcd13a7 09ec98f 9cbc665 bcd13a7 09ec98f bcd13a7 9cbc665 bcd13a7 9cbc665 bcd13a7 9cbc665 bcd13a7 09ec98f bcd13a7 9cbc665 bcd13a7 9cbc665 09ec98f a4dd486 bcd13a7 09ec98f b8d0a87 9cbc665 a4dd486 bcd13a7 09ec98f b8d0a87 9cbc665 a4dd486 bcd13a7 09ec98f a4dd486 bcd13a7 09ec98f 9cbc665 a4dd486 bcd13a7 9cbc665 bcd13a7 09ec98f a4dd486 9cbc665 09ec98f b8d0a87 09ec98f b8d0a87 09ec98f bcd13a7 | 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 | """Constrained CE over answer codes (matches serving softmax).
Full-vocab next-token CE left relative Y/N logits dominated by the
base prior (eval collapse: all Noul → Y). Scoring only code tokens
aligns train with infer.
Also: Noul consistency (P(yes|Q)+P(yes|¬Q)→1), optional binary
head CE on last-token hidden, and Score ordinal EMD (levels ordered).
"""
from __future__ import annotations
from typing import Any
import torch
import torch.nn.functional as F
def _answer_pos(shift_labels: torch.Tensor) -> int | None:
pos = (shift_labels != -100).nonzero(as_tuple=True)[0]
if len(pos) == 0:
return None
return int(pos[0].item())
def code_token_loss(
logits: torch.Tensor,
labels: torch.Tensor,
code_token_ids: torch.Tensor,
n_codes: torch.Tensor,
soft_labels: torch.Tensor,
*,
use_soft: bool,
sample_weight: torch.Tensor | None = None,
force_soft_mask: torch.Tensor | None = None,
) -> torch.Tensor:
"""Softmax CE / soft-KL over allowed code ids at the answer position.
``logits``: [B, T, V]; ``labels``: [B, T] with -100 mask (gold at
the answer index). HF shift: predict token t from logits[:, t-1].
Soft path is KL(teacher‖student) up to H(teacher).
``force_soft_mask``: per-row override to soft-KL (Score rows).
"""
shift_logits = logits[:, :-1, :].contiguous()
shift_labels = labels[:, 1:].contiguous()
batch = shift_logits.size(0)
losses: list[torch.Tensor] = []
weights: list[torch.Tensor] = []
for b in range(batch):
p = _answer_pos(shift_labels[b])
if p is None:
continue
k = int(n_codes[b].item())
if k < 1:
continue
ids = code_token_ids[b, :k].long()
code_logits = shift_logits[b, p, ids]
row_soft = use_soft
if force_soft_mask is not None and int(force_soft_mask[b].item()) > 0:
row_soft = True
if row_soft:
target = soft_labels[b, :k].float()
target = target / target.sum().clamp_min(1e-8)
log_p = F.log_softmax(code_logits, dim=-1)
loss_b = -(target * log_p).sum()
else:
gold_idx = int(soft_labels[b, :k].argmax().item())
loss_b = F.cross_entropy(
code_logits.unsqueeze(0),
torch.tensor(
[gold_idx],
device=code_logits.device,
dtype=torch.long,
),
)
losses.append(loss_b)
if sample_weight is not None:
weights.append(sample_weight[b].float())
else:
weights.append(
torch.ones((), device=code_logits.device)
)
if not losses:
return logits.sum() * 0.0
stacked = torch.stack(losses)
w = torch.stack(weights)
return (stacked * w).sum() / w.sum().clamp_min(1e-8)
def ordinal_emd_loss(
logits: torch.Tensor,
labels: torch.Tensor,
code_token_ids: torch.Tensor,
n_codes: torch.Tensor,
soft_labels: torch.Tensor,
is_score: torch.Tensor,
*,
sample_weight: torch.Tensor | None = None,
) -> torch.Tensor:
"""1D Earth-mover distance on ordered Score levels.
Penalizes mass 3 levels off more than 1: W1 = Σ|CDF_p − CDF_t|.
Only rows with ``is_score>0`` contribute.
"""
shift_logits = logits[:, :-1, :].contiguous()
shift_labels = labels[:, 1:].contiguous()
batch = shift_logits.size(0)
losses: list[torch.Tensor] = []
weights: list[torch.Tensor] = []
for b in range(batch):
if int(is_score[b].item()) <= 0:
continue
p = _answer_pos(shift_labels[b])
if p is None:
continue
k = int(n_codes[b].item())
if k < 2:
continue
ids = code_token_ids[b, :k].long()
code_logits = shift_logits[b, p, ids]
pred = F.softmax(code_logits.float(), dim=-1)
target = soft_labels[b, :k].float()
target = target / target.sum().clamp_min(1e-8)
# Also expected |level−gold| so hard-off-by-3 is explicit.
levels = torch.arange(
k, device=code_logits.device, dtype=pred.dtype
)
gold = float((target * levels).sum().item())
mae = (pred * (levels - gold).abs()).sum()
emd = (pred.cumsum(0) - target.cumsum(0)).abs().sum()
losses.append(emd + mae)
if sample_weight is not None:
weights.append(sample_weight[b].float())
else:
weights.append(
torch.ones((), device=code_logits.device)
)
if not losses:
return logits.sum() * 0.0
stacked = torch.stack(losses)
w = torch.stack(weights)
return (stacked * w).sum() / w.sum().clamp_min(1e-8)
def code_yes_probs(
logits: torch.Tensor,
labels: torch.Tensor,
code_token_ids: torch.Tensor,
n_codes: torch.Tensor,
yes_idx: torch.Tensor,
) -> torch.Tensor:
"""P(yes) from constrained softmax; -1 where not applicable."""
shift_logits = logits[:, :-1, :].contiguous()
shift_labels = labels[:, 1:].contiguous()
batch = shift_logits.size(0)
out = torch.full(
(batch,), -1.0, device=logits.device, dtype=logits.dtype
)
for b in range(batch):
yi = int(yes_idx[b].item()) if yes_idx is not None else -1
if yi < 0:
continue
p = _answer_pos(shift_labels[b])
if p is None:
continue
k = int(n_codes[b].item())
if k < 2 or yi >= k:
continue
ids = code_token_ids[b, :k].long()
probs = F.softmax(shift_logits[b, p, ids].float(), dim=-1)
out[b] = probs[yi]
return out
def consistency_loss(
p_yes_q: torch.Tensor,
p_yes_neg: torch.Tensor,
has_neg: torch.Tensor,
) -> torch.Tensor:
"""(P(yes|Q) + P(yes|¬Q) − 1)^2 on rows with a negation twin."""
mask = (has_neg > 0) & (p_yes_q >= 0) & (p_yes_neg >= 0)
if not mask.any():
return p_yes_q.sum() * 0.0
diff = p_yes_q[mask] + p_yes_neg[mask] - 1.0
return (diff * diff).mean()
def trace_token_loss(
logits: torch.Tensor,
trace_labels: torch.Tensor,
) -> torch.Tensor:
"""Full-vocab CE on rationale tokens (Path A L_trace).
``trace_labels``: [B, T] with -100 mask; only rationale positions
supervised. HF shift: predict token t from logits[:, t-1].
"""
shift_logits = logits[:, :-1, :].contiguous()
shift_labels = trace_labels[:, 1:].contiguous()
if (shift_labels != -100).sum() == 0:
return logits.sum() * 0.0
return F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
ignore_index=-100,
)
class CodeTokenTrainer:
"""Lazy wrapper — real subclass built on first instantiate.
Avoids importing transformers at module import time (unit tests).
"""
def __new__(cls, *args: Any, **kwargs: Any):
return _build_trainer_class()(*args, **kwargs)
_TRAINER_CLS: type | None = None
def _build_trainer_class() -> type:
global _TRAINER_CLS
if _TRAINER_CLS is not None:
return _TRAINER_CLS
from transformers import Trainer
class _CodeTokenTrainer(Trainer):
"""Trainer that CE-supervises only the allowed answer codes."""
def __init__(
self,
*args: Any,
use_soft_ce: bool = False,
train_sampler: Any | None = None,
consistency_weight: float = 0.0,
binary_head: Any | None = None,
binary_head_weight: float = 0.0,
binary_head_balanced: bool = False,
score_ordinal_weight: float = 0.0,
trace_loss_weight: float = 0.0,
**kwargs: Any,
) -> None:
self._custom_train_sampler = train_sampler
super().__init__(*args, **kwargs)
self.use_soft_ce = bool(use_soft_ce)
self.consistency_weight = float(consistency_weight)
self.binary_head = binary_head
self.binary_head_weight = float(binary_head_weight)
self.binary_head_balanced = bool(binary_head_balanced)
self.score_ordinal_weight = float(score_ordinal_weight)
self.trace_loss_weight = float(trace_loss_weight)
def _get_train_sampler(self, train_dataset: Any = None):
if self._custom_train_sampler is not None:
return self._custom_train_sampler
try:
return super()._get_train_sampler(train_dataset)
except TypeError:
return super()._get_train_sampler()
def compute_loss(
self,
model: Any,
inputs: dict[str, Any],
return_outputs: bool = False,
num_items_in_batch: int | None = None,
):
code_token_ids = inputs.pop("code_token_ids")
n_codes = inputs.pop("n_codes")
soft_labels = inputs.pop("soft_labels")
sample_weight = inputs.pop("sample_weight", None)
yes_idx = inputs.pop("yes_idx", None)
has_neg = inputs.pop("has_neg", None)
neg_input_ids = inputs.pop("neg_input_ids", None)
neg_attention = inputs.pop("neg_attention_mask", None)
neg_labels = inputs.pop("neg_labels", None)
neg_soft = inputs.pop("neg_soft_labels", None)
trace_labels = inputs.pop("trace_labels", None)
inputs.pop("label_id", None)
inputs.pop("has_soft", None)
is_noul = inputs.pop("is_noul", None)
is_score = inputs.pop("is_score", None)
force_soft = inputs.pop("force_soft", None)
inputs.pop("has_trace", None)
labels = inputs.get("labels")
outputs = model(**inputs)
loss = code_token_loss(
outputs.logits,
labels,
code_token_ids,
n_codes,
soft_labels,
use_soft=self.use_soft_ce,
sample_weight=sample_weight,
force_soft_mask=force_soft,
)
# Path A: λ_t L_trace on short rationale tokens.
if (
self.trace_loss_weight > 0
and trace_labels is not None
):
t_loss = trace_token_loss(
outputs.logits, trace_labels
)
loss = loss + self.trace_loss_weight * t_loss
# Score ordinal: EMD + expected |level−teacher| (off-by-3 > 1).
if (
self.score_ordinal_weight > 0
and is_score is not None
and (is_score > 0).any()
):
ord_loss = ordinal_emd_loss(
outputs.logits,
labels,
code_token_ids,
n_codes,
soft_labels,
is_score,
sample_weight=sample_weight,
)
loss = loss + self.score_ordinal_weight * ord_loss
# Negation consistency + CE on ¬Q (dual forward).
if (
self.consistency_weight > 0
and neg_input_ids is not None
and has_neg is not None
and (has_neg > 0).any()
):
neg_out = model(
input_ids=neg_input_ids,
attention_mask=neg_attention,
)
loss = loss + code_token_loss(
neg_out.logits,
neg_labels,
code_token_ids,
n_codes,
neg_soft if neg_soft is not None else soft_labels,
use_soft=self.use_soft_ce,
sample_weight=sample_weight,
)
if yes_idx is not None:
p_q = code_yes_probs(
outputs.logits,
labels,
code_token_ids,
n_codes,
yes_idx,
)
p_n = code_yes_probs(
neg_out.logits,
neg_labels,
code_token_ids,
n_codes,
yes_idx,
)
cons = consistency_loss(p_q, p_n, has_neg)
loss = loss + self.consistency_weight * cons
# Optional dedicated binary head on last non-pad hidden.
if (
self.binary_head is not None
and self.binary_head_weight > 0
and is_noul is not None
and yes_idx is not None
):
hidden = outputs.hidden_states
if hidden is None:
# Re-forward with hidden states if model omitted them.
outputs2 = model(
**{
k: v
for k, v in inputs.items()
if k
in ("input_ids", "attention_mask")
},
output_hidden_states=True,
)
hidden = outputs2.hidden_states
if hidden is not None:
h = hidden[-1]
attn = inputs["attention_mask"]
# Last non-pad index per row.
lengths = attn.long().sum(dim=1) - 1
b_idx = torch.arange(h.size(0), device=h.device)
pooled = h[b_idx, lengths.clamp_min(0)]
logits2 = self.binary_head(pooled)
# Target: 1 if gold is yes-code else 0.
gold = soft_labels.argmax(dim=-1)
target = (gold == yes_idx.clamp_min(0)).long()
# Only noul rows.
mask = is_noul > 0
if mask.any():
logits_m = logits2[mask]
target_m = target[mask]
weight = None
if self.binary_head_balanced:
# Inverse-frequency class weights so Y/N
# each contribute ~half the binary CE
# (train aux; LM Y/N remains serving).
n0 = (target_m == 0).sum().clamp_min(1)
n1 = (target_m == 1).sum().clamp_min(1)
tot = (n0 + n1).to(logits_m.dtype)
weight = torch.stack(
[
tot / (2.0 * n0.to(logits_m.dtype)),
tot / (2.0 * n1.to(logits_m.dtype)),
]
)
bce = F.cross_entropy(
logits_m, target_m, weight=weight
)
loss = (
loss + self.binary_head_weight * bce
)
return (loss, outputs) if return_outputs else loss
_TRAINER_CLS = _CodeTokenTrainer
return _TRAINER_CLS
|