Spaces:
Running
Running
File size: 55,368 Bytes
11463c3 27612e1 11463c3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 | """
GreesyGPT β Content-moderation language model with KV caching and dropout.
Production-ready implementation featuring:
β’ KV caching for O(1) per-token inference (instead of recomputing full sequence)
β’ Configurable dropout for regularisation during training
β’ Centralised ModelConfig dataclass for all hyperparameters
β’ Structured logging via the stdlib ``logging`` module
β’ Type annotations throughout
"""
from __future__ import annotations
import contextlib
import json
import logging
import re
from dataclasses import dataclass
from enum import Enum
from pathlib import Path
from typing import Any, Optional, cast
import tiktoken
import torch
import torch.nn as nn
from torch.nn import functional as F
from torch.nn.utils.rnn import pad_sequence
from torch.utils.data import DataLoader, Dataset
from tqdm import tqdm
# βββββββββββββββββββββββββββββββββββββββββββββ
# Logging
# βββββββββββββββββββββββββββββββββββββββββββββ
logger = logging.getLogger("greesygpt")
if not logger.handlers:
_handler = logging.StreamHandler()
_handler.setFormatter(logging.Formatter("[%(levelname)s] %(name)s: %(message)s"))
logger.addHandler(_handler)
logger.setLevel(logging.INFO)
# βββββββββββββββββββββββββββββββββββββββββββββ
# Model Configuration
# βββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class ModelConfig:
"""Centralised hyperparameter store for GreesyGPT."""
vocab_size: int = 200032
context_len: int = 12_000
n_embd: int = 768
n_head: int = 12
n_layer: int = 12
# Dropout rates (set to 0.0 at inference via model.eval(); typical training values 0.1β0.2)
attn_dropout: float = 0.1
resid_dropout: float = 0.1
embd_dropout: float = 0.1
mlp_dropout: float = 0.1
@property
def head_dim(self) -> int:
assert self.n_embd % self.n_head == 0, "n_embd must be divisible by n_head"
return self.n_embd // self.n_head
# Legacy constants (kept for backward compatibility; prefer ModelConfig)
DEFAULT_CONFIG = ModelConfig()
VOCAB_SIZE = DEFAULT_CONFIG.vocab_size
CONTEXT_LEN = DEFAULT_CONFIG.context_len
N_EMBD = DEFAULT_CONFIG.n_embd
N_HEAD = DEFAULT_CONFIG.n_head
N_LAYER = DEFAULT_CONFIG.n_layer
# βββββββββββββββββββββββββββββββββββββββββββββ
# Device (MPS β CUDA β CPU, in priority order)
# βββββββββββββββββββββββββββββββββββββββββββββ
def _select_device() -> torch.device:
if torch.backends.mps.is_available() and torch.backends.mps.is_built():
return torch.device("mps")
if torch.cuda.is_available():
return torch.device("cuda")
return torch.device("cpu")
DEVICE: torch.device = _select_device()
def _autocast_ctx():
"""
Return the appropriate mixed-precision context for the active device.
β’ MPS β float16 (bfloat16 not yet supported by MPS kernel)
β’ CUDA β bfloat16 (preferred for training stability)
β’ CPU β no-op (autocast on CPU is rarely beneficial)
"""
if DEVICE.type == "mps":
return torch.autocast(device_type="mps", dtype=torch.float16)
if DEVICE.type == "cuda":
return torch.autocast(device_type="cuda", dtype=torch.bfloat16)
return contextlib.nullcontext()
# βββββββββββββββββββββββββββββββββββββββββββββ
# Special-token registry
# 199999 <|endoftext|> (native to o200k_base)
# 200019 <think> (first free slot after o200k_base's 200019 mergeable ranks)
# 200020 </think>
# 200021 <|system|>
# 200022 </|system|>
# 200023 <|user|>
# 200024 </|user|>
# 200025 <|assistant|> (turn closed by <|endoftext|>, no explicit close tag)
# βββββββββββββββββββββββββββββββββββββββββββββ
SPECIAL_TOKENS: dict[str, int] = {
"<|endoftext|>": 199999,
"<think>": 200019,
"</think>": 200020,
"<|system|>": 200021,
"</|system|>": 200022,
"<|user|>": 200023,
"</|user|>": 200024,
"<|assistant|>": 200025,
}
# βββββββββββββββββββββββββββββββββββββββββββββ
# Tokenizer
# βββββββββββββββββββββββββββββββββββββββββββββ
def get_tokenizer() -> tiktoken.Encoding:
base = tiktoken.get_encoding("o200k_base")
return tiktoken.Encoding(
name="greesy_gpt",
pat_str=base._pat_str,
mergeable_ranks=base._mergeable_ranks,
special_tokens={**base._special_tokens, **SPECIAL_TOKENS},
)
tokenizer = get_tokenizer()
# Convenience single-token IDs used throughout
def _tid(s: str) -> int:
return tokenizer.encode(s, allowed_special="all")[0]
TOK_EOT = _tid("<|endoftext|>")
TOK_THINK_OPEN = _tid("<think>")
TOK_THINK_CLOSE = _tid("</think>")
TOK_SYSTEM = _tid("<|system|>")
TOK_ESYSTEM = _tid("</|system|>")
TOK_USER = _tid("<|user|>")
TOK_EUSER = _tid("</|user|>")
TOK_ASSISTANT = _tid("<|assistant|>")
# βββββββββββββββββββββββββββββββββββββββββββββ
# Chat Template
# βββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class Message:
"""A single turn in a conversation."""
role: str # "system" | "user" | "assistant"
content: str
class ChatTemplate:
"""
Serialises a ``list[Message]`` into GreesyGPT's role-delimited wire format:
<|system|>
{system content}
</|system|>
<|user|>
{user content}
</|user|>
<|assistant|>
<think>
{reasoning}
</think>
{verdict}<|endoftext|>
For training, only the *assistant completion* tokens (everything after
``<|assistant|>\\n``) contribute to the loss; all other tokens are masked
with ``-100`` in the labels tensor.
For multi-turn conversations append additional user/assistant message
pairs β the template handles them in sequence.
"""
_OPEN = {"system": "<|system|>", "user": "<|user|>", "assistant": "<|assistant|>"}
_CLOSE = {"system": "</|system|>", "user": "</|user|>", "assistant": ""} # EOT closes assistant
# ββ Rendering βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@classmethod
def render(cls, messages: list[Message], add_generation_prompt: bool = False) -> str:
"""
Render a full conversation to a single string.
Parameters
----------
messages:
Ordered ``Message`` list (system, then alternating user/assistant).
add_generation_prompt:
Append ``<|assistant|>\\n<think>\\n`` so the model can continue
from the correct position during inference.
"""
parts: list[str] = []
for msg in messages:
open_tag = cls._OPEN[msg.role]
close_tag = cls._CLOSE[msg.role]
if close_tag:
parts.append(f"{open_tag}\n{msg.content}\n{close_tag}\n")
else:
# assistant: content already contains <think>β¦</think>+verdict+EOT
parts.append(f"{open_tag}\n{msg.content}")
if add_generation_prompt:
parts.append("<|assistant|>\n<think>\n")
return "".join(parts)
@staticmethod
def render_assistant_content(reasoning: str, verdict: str) -> str:
"""Build the assistant ``content`` field from reasoning + verdict."""
return f"<think>\n{reasoning}\n</think>\n{verdict}<|endoftext|>"
# ββ Tokenisation with per-role label masking ββββββββββββββββββββββββββββββ
@classmethod
def tokenize(
cls,
messages: list[Message],
enc: tiktoken.Encoding,
max_length: int = 12288,
) -> dict[str, torch.Tensor]:
"""
Tokenise a conversation and return ``input_ids`` + ``labels``.
All system and user tokens are masked (``-100``).
Only assistant completion tokens are trained.
Supports multi-turn: each assistant turn is unmasked individually.
"""
input_ids: list[int] = []
labels: list[int] = []
for msg in messages:
open_tag = cls._OPEN[msg.role]
close_tag = cls._CLOSE[msg.role]
if msg.role == "assistant":
# Role header β masked
header_toks = enc.encode(f"{open_tag}\n", allowed_special="all")
input_ids.extend(header_toks)
labels.extend([-100] * len(header_toks))
# Completion β trained
comp_toks = enc.encode(msg.content, allowed_special="all")
input_ids.extend(comp_toks)
labels.extend(comp_toks)
else:
text = (
f"{open_tag}\n{msg.content}\n{close_tag}\n"
if close_tag
else f"{open_tag}\n{msg.content}"
)
toks = enc.encode(text, allowed_special="all")
input_ids.extend(toks)
labels.extend([-100] * len(toks))
# Truncate to max_length
input_ids = input_ids[:max_length]
labels = labels[:max_length]
return {
"input_ids": torch.tensor(input_ids, dtype=torch.long),
"labels": torch.tensor(labels, dtype=torch.long),
}
# ββ Inference helper ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@classmethod
def build_inference_prompt(cls, user_message: str, system_prompt: str = "") -> str:
"""Return the prompt string to feed the model at inference time."""
msgs: list[Message] = []
if system_prompt:
msgs.append(Message("system", system_prompt))
msgs.append(Message("user", user_message))
return cls.render(msgs, add_generation_prompt=True)
# βββββββββββββββββββββββββββββββββββββββββββββ
# Output Formats & Markdown helpers
# βββββββββββββββββββββββββββββββββββββββββββββ
class OutputFormat(Enum):
"""
MARKDOWN β raw model output; headings, bold, and lists preserved.
PLAIN β markdown stripped to clean prose.
JSON β structured dict with label, severity, confidence, and summary.
"""
MARKDOWN = "markdown"
PLAIN = "plain"
JSON = "json"
# Severity scale used by the JSON formatter
_LABEL_SEVERITY: dict[str, int] = {
"SAFE": 0,
"SPAM": 1,
"MISINFORMATION": 2,
"HARASSMENT": 3,
"HATE_SPEECH": 4,
"CRISIS_REFERRAL": 5,
"UNSAFE": 6,
}
# Ordered substitutions for stripping Markdown syntax
_MD_STRIP: list[tuple[re.Pattern[str], str]] = [
(re.compile(r"#{1,6}\s*"), ""), # headings
(re.compile(r"\*\*(.+?)\*\*"), r"\1"), # bold
(re.compile(r"\*(.+?)\*"), r"\1"), # italic
(re.compile(r"`{1,3}(.+?)`{1,3}", re.DOTALL), r"\1"), # inline/fenced code
(re.compile(r"^\s*[-*+]\s+", re.M), "β’ "), # unordered list
(re.compile(r"^\s*\d+\.\s+", re.M), ""), # ordered list numbers
(re.compile(r"\[(.+?)\]\(.+?\)"), r"\1"), # links
(re.compile(r"^\s*>\s?", re.M), ""), # blockquotes
(re.compile(r"-{3,}|={3,}|\*{3,}"), ""), # horizontal rules
(re.compile(r"\n{3,}"), "\n\n"),# excess blank lines
]
def strip_markdown(text: str) -> str:
"""Remove common Markdown syntax and return clean prose."""
for pattern, repl in _MD_STRIP:
text = pattern.sub(repl, text)
return text.strip()
def extract_verdict_label(verdict_text: str) -> str:
"""Pull the label keyword from a verdict block. Falls back to 'UNKNOWN'."""
upper = verdict_text.upper()
for label in _LABEL_SEVERITY:
if label in upper:
return label
return "UNKNOWN"
def format_output(result: dict[str, Any], fmt: OutputFormat = OutputFormat.MARKDOWN) -> "str | dict[str, Any]":
"""
Post-process a ``generate_moderation`` result.
Returns ``str`` for MARKDOWN/PLAIN, ``dict`` for JSON.
"""
verdict = result.get("verdict", "")
thinking = result.get("thinking") or ""
mode = result.get("mode")
if fmt == OutputFormat.PLAIN:
return strip_markdown(verdict)
if fmt == OutputFormat.JSON:
label = extract_verdict_label(verdict)
sev = _LABEL_SEVERITY.get(label, -1)
conf = "high" if sev <= 1 else ("medium" if sev <= 3 else "low")
return {
"verdict": label,
"severity": sev,
"confidence_hint": conf,
"reasoning_mode": mode.value if mode else None,
"thinking_summary": thinking[:300].strip(),
"full_verdict": verdict,
}
return verdict # MARKDOWN: return as-is
# βββββββββββββββββββββββββββββββββββββββββββββ
# Reasoning Modes
# βββββββββββββββββββββββββββββββββββββββββββββ
class ReasoningMode(Enum):
"""
Controls how much thinking the model does before emitting a verdict.
NONE β minimal chain-of-thought; fastest, best for obvious cases.
LOW β balanced reasoning; good general-purpose default.
MEDIUM β extended deliberation; best for nuanced / borderline content.
HIGH β maximum tokens + lower temperature; use for high-stakes review.
"""
NONE = "none"
LOW = "low"
MEDIUM = "medium"
HIGH = "high"
# Injected into every system prompt to teach the model Markdown output style
_MARKDOWN_INSTRUCTION = (
"Format your entire response in **Markdown**:\n"
"- Use `##` headings to separate reasoning sections "
"(e.g. `## Intent`, `## Evidence`, `## Harm Potential`, `## Edge Cases`).\n"
"- Use `**bold**` to highlight key terms, policy labels, and the final verdict.\n"
"- Use bullet lists (`-`) for evidence points.\n"
"- Use `>` blockquotes for direct quotes from the message under review.\n"
"- End with a `## Verdict` section containing **only** the label on its own line "
"(e.g. `**SAFE**`, `**HARASSMENT**`).\n"
)
@dataclass
class ReasoningConfig:
max_think_tokens: int # token budget for <think>β¦</think>
max_total_tokens: int # total generation budget (think + verdict)
temperature: float
top_k: int
system_prompt: str # full system-turn content
def total_budget(self) -> int:
return self.max_total_tokens
REASONING_CONFIGS: dict[ReasoningMode, ReasoningConfig] = {
ReasoningMode.NONE: ReasoningConfig(
max_think_tokens=200,
max_total_tokens=812,
temperature=0.1,
top_k=50,
system_prompt=(
"You are a fast content moderator. "
"Think briefly, then give a clear verdict.\n\n"
+ _MARKDOWN_INSTRUCTION
),
),
ReasoningMode.LOW: ReasoningConfig(
max_think_tokens=512,
max_total_tokens=1200,
temperature=0.7,
top_k=40,
system_prompt=(
"You are a careful content moderator. "
"Reason step-by-step, then issue a structured Markdown verdict.\n\n"
+ _MARKDOWN_INSTRUCTION
),
),
ReasoningMode.MEDIUM: ReasoningConfig(
max_think_tokens=1536,
max_total_tokens=2048,
temperature=0.6,
top_k=30,
system_prompt=(
"You are a thorough content moderator. "
"Analyse intent, context, potential harm, and edge cases "
"before issuing a detailed Markdown verdict.\n\n"
+ _MARKDOWN_INSTRUCTION
),
),
ReasoningMode.HIGH: ReasoningConfig(
max_think_tokens=3072,
max_total_tokens=4096,
temperature=0.4,
top_k=20,
system_prompt=(
"You are an expert content safety reviewer. "
"Examine the message from every relevant angleβlegal, ethical, "
"contextual, and platform-policyβand produce a comprehensive "
"Markdown safety report.\n\n"
+ _MARKDOWN_INSTRUCTION
),
),
}
def describe_reasoning_modes() -> str:
lines = ["Available Reasoning Modes\n" + "=" * 44]
for mode, cfg in REASONING_CONFIGS.items():
lines.append(
f" {mode.value:10s} thinkβ€{cfg.max_think_tokens:4d} tokens | "
f"totalβ€{cfg.max_total_tokens:4d} tokens | "
f"temp={cfg.temperature} | top_k={cfg.top_k}"
)
return "\n".join(lines)
DATASET_JSON_PATH = Path(__file__).with_name("dataset.json")
# βββββββββββββββββββββββββββββββββββββββββββββ
# KV Cache
# βββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class KVCache:
"""
Per-layer key/value cache for autoregressive generation.
Stores tensors of shape ``[B, n_head, T_cached, head_dim]``.
Grows incrementally as new tokens are generated.
"""
key: Optional[torch.Tensor] = None
value: Optional[torch.Tensor] = None
@property
def seq_len(self) -> int:
"""Number of tokens currently cached."""
if self.key is None:
return 0
return self.key.shape[2]
def update(
self, new_key: torch.Tensor, new_value: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Append new K/V slices and return the full accumulated tensors.
Parameters
----------
new_key, new_value : ``[B, n_head, T_new, head_dim]``
Returns
-------
(full_key, full_value) each ``[B, n_head, T_total, head_dim]``
"""
if self.key is None or self.value is None:
self.key = new_key
self.value = new_value
else:
assert self.key is not None and self.value is not None
self.key = torch.cat((self.key, new_key), dim=2)
self.value = torch.cat((self.value, new_value), dim=2)
return cast(torch.Tensor, self.key), cast(torch.Tensor, self.value)
def clear(self) -> None:
self.key = None
self.value = None
# Type alias: one KVCache per layer
LayerCaches = list[KVCache]
def make_kv_caches(n_layers: int) -> LayerCaches:
"""Create a fresh list of empty KV caches, one per transformer layer."""
return [KVCache() for _ in range(n_layers)]
# βββββββββββββββββββββββββββββββββββββββββββββ
# RoPE (supports position offset for KV cache)
# βββββββββββββββββββββββββββββββββββββββββββββ
class RoPE(nn.Module):
def __init__(self, head_dim: int, max_seq_len: int = CONTEXT_LEN):
super().__init__()
inv_freq = 1.0 / (10000.0 ** (torch.arange(0, head_dim, 2).float() / head_dim))
self.register_buffer("inv_freq", inv_freq)
def forward(
self, seq_len: int, device: torch.device, offset: int = 0
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Compute cos/sin embeddings for positions ``[offset, offset+seq_len)``.
Parameters
----------
seq_len : number of new positions to compute
device : target device
offset : starting position index (= number of previously cached tokens)
"""
inv_freq = cast(torch.Tensor, self.inv_freq)
t: torch.Tensor = torch.arange(offset, offset + seq_len, device=device, dtype=inv_freq.dtype)
freqs = torch.einsum("i,j->ij", t, inv_freq)
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos()[None, :, None, :], emb.sin()[None, :, None, :]
def rotate_half(x: torch.Tensor) -> torch.Tensor:
half = x.shape[-1] // 2
return torch.cat((-x[..., half:], x[..., :half]), dim=-1)
def apply_rope(q, k, cos, sin):
q = (q * cos) + (rotate_half(q) * sin)
k = (k * cos) + (rotate_half(k) * sin)
return q, k
# βββββββββββββββββββββββββββββββββββββββββββββ
# Transformer Block (with KV cache + dropout)
# βββββββββββββββββββββββββββββββββββββββββββββ
class GreesyBlock(nn.Module):
def __init__(self, config: ModelConfig):
super().__init__()
n_embd = config.n_embd
self.n_head = config.n_head
self.head_dim = config.head_dim
self.ln1 = nn.LayerNorm(n_embd)
self.ln2 = nn.LayerNorm(n_embd)
self.qkv = nn.Linear(n_embd, 3 * n_embd, bias=False)
self.out_proj = nn.Linear(n_embd, n_embd, bias=False)
self.rope = RoPE(self.head_dim, max_seq_len=config.context_len)
# Dropout layers
self.attn_dropout = nn.Dropout(config.attn_dropout)
self.resid_dropout = nn.Dropout(config.resid_dropout)
self.mlp = nn.Sequential(
nn.Linear(n_embd, 4 * n_embd),
nn.GELU(),
nn.Dropout(config.mlp_dropout),
nn.Linear(4 * n_embd, n_embd),
nn.Dropout(config.resid_dropout),
)
def forward(
self,
x: torch.Tensor,
kv_cache: Optional[KVCache] = None,
) -> torch.Tensor:
"""
Forward pass with optional KV caching.
Parameters
----------
x : ``[B, T, C]`` β input embeddings (full sequence or single new token)
kv_cache : if provided, keys/values are appended to the cache and the
full cached K/V are used for attention, enabling O(1) per-token
inference instead of O(T).
"""
B, T, C = x.shape
norm_x = self.ln1(x)
qkv = self.qkv(norm_x).reshape(B, T, 3, self.n_head, self.head_dim)
q, k, v = qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2]
# Position offset = number of tokens already in the cache
offset = kv_cache.seq_len if kv_cache is not None else 0
cos, sin = self.rope(T, x.device, offset=offset)
q, k = apply_rope(q, k, cos, sin)
# [B, T, n_head, head_dim] β [B, n_head, T, head_dim]
q = q.transpose(1, 2)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
# Update KV cache (if provided) and use full cached K/V for attention
if kv_cache is not None:
k, v = kv_cache.update(k, v)
# Causal mask is only needed during training/prefill (offset==0).
# During cached single-token generation (offset>0, T_q=1) every
# cached position is visible, so is_causal=False is correct.
is_causal = kv_cache is None or offset == 0
attn_out = F.scaled_dot_product_attention(
q, k, v,
is_causal=is_causal,
dropout_p=self.attn_dropout.p if self.training else 0.0,
)
# [B, n_head, T, head_dim] β [B, T, C]
attn_out = attn_out.transpose(1, 2).reshape(B, T, C)
x = x + self.resid_dropout(self.out_proj(attn_out))
x = x + self.mlp(self.ln2(x))
return x
# βββββββββββββββββββββββββββββββββββββββββββββ
# Model (config-driven, with embedding dropout + KV cache support)
# βββββββββββββββββββββββββββββββββββββββββββββ
class GreesyGPT(nn.Module):
def __init__(self, config: Optional[ModelConfig] = None):
super().__init__()
self.config = config or DEFAULT_CONFIG
c = self.config
self.embd_dropout = nn.Dropout(c.embd_dropout)
self.tok_emb = nn.Embedding(c.vocab_size, c.n_embd)
self.blocks = nn.ModuleList([GreesyBlock(c) for _ in range(c.n_layer)])
self.ln_f = nn.LayerNorm(c.n_embd)
self.head = nn.Linear(c.n_embd, c.vocab_size, bias=False)
self.tok_emb.weight = self.head.weight # weight tying
n_params = sum(p.numel() for p in self.parameters())
logger.info(
"GreesyGPT initialised β %.2fM params, %d layers, %d heads, "
"ctx=%d, dropout=(attn=%.2f, resid=%.2f, embd=%.2f, mlp=%.2f)",
n_params / 1e6, c.n_layer, c.n_head, c.context_len,
c.attn_dropout, c.resid_dropout, c.embd_dropout, c.mlp_dropout,
)
def forward(
self,
idx: torch.Tensor,
kv_caches: Optional[LayerCaches] = None,
) -> torch.Tensor:
"""
Forward pass.
Parameters
----------
idx : ``[B, T]`` token indices
kv_caches : optional list of ``KVCache`` (one per layer). When
provided, enables incremental decoding.
Returns
-------
logits : ``[B, T, vocab_size]``
"""
x = self.embd_dropout(self.tok_emb(idx))
for i, block in enumerate(self.blocks):
cache = kv_caches[i] if kv_caches is not None else None
x = block(x, kv_cache=cache)
return self.head(self.ln_f(x))
# βββββββββββββββββββββββββββββββββββββββββββββ
# Generation (KV-cached, chat-template + reasoning-mode aware)
# βββββββββββββββββββββββββββββββββββββββββββββ
@torch.inference_mode()
def generate_moderation(
model: GreesyGPT,
prompt: str,
mode: ReasoningMode = ReasoningMode.LOW,
output_format: OutputFormat = OutputFormat.MARKDOWN,
# Optional per-call overrides (take priority over mode defaults)
max_tokens: Optional[int] = None,
temp: Optional[float] = None,
top_k: Optional[int] = None,
use_kv_cache: bool = True,
) -> dict[str, Any]:
"""
Run moderation inference via the chat template with KV caching.
Parameters
----------
model : trained GreesyGPT
prompt : raw user message to moderate
mode : ReasoningMode (controls budget, temperature, system prompt)
output_format : post-processing format for the verdict
max_tokens : overrides mode's ``max_total_tokens``
temp : overrides mode's ``temperature``
top_k : overrides mode's ``top_k``
use_kv_cache : if True (default), use KV caching for efficient generation
Returns
-------
dict with keys
full_text raw decoded sequence (includes special tokens)
thinking <think>β¦</think> block content (str | None)
verdict raw Markdown verdict string
verdict_fmt post-processed verdict (str or dict depending on output_format)
mode ReasoningMode used
output_format OutputFormat used
"""
cfg = REASONING_CONFIGS[mode]
_max = max_tokens if max_tokens is not None else cfg.max_total_tokens
_temp = temp if temp is not None else cfg.temperature
_topk = top_k if top_k is not None else cfg.top_k
model.eval()
input_str = ChatTemplate.build_inference_prompt(
user_message=prompt,
system_prompt=cfg.system_prompt,
)
tokens = tokenizer.encode(input_str, allowed_special="all")
context_len = model.config.context_len
think_tokens = 0
think_closed = False
if use_kv_cache:
# ββ KV-cached generation ββββββββββββββββββββββββββββββββββββββββββ
kv_caches = make_kv_caches(model.config.n_layer)
idx = torch.tensor([tokens], device=DEVICE)
# Truncate prompt if it exceeds context length
if idx.shape[1] > context_len:
idx = idx[:, -context_len:]
logger.warning("Prompt truncated to context_len=%d tokens", context_len)
# Prefill: process the entire prompt in one forward pass
logits = model(idx, kv_caches=kv_caches)
generated_ids: list[int] = []
for _ in range(_max):
scaled_logits = logits[:, -1, :] / _temp
if not think_closed and think_tokens >= cfg.max_think_tokens:
next_id = TOK_THINK_CLOSE
next_tok = torch.tensor([[next_id]], device=DEVICE)
else:
v, _ = torch.topk(scaled_logits, _topk)
scaled_logits[scaled_logits < v[:, [-1]]] = -float("Inf")
probs = F.softmax(scaled_logits, dim=-1)
next_tok = torch.multinomial(probs, num_samples=1)
next_id = int(next_tok.item())
generated_ids.append(next_id)
if not think_closed:
if next_id == TOK_THINK_CLOSE:
think_closed = True
else:
think_tokens += 1
if next_id == TOK_EOT:
break
# Check context length limit
if kv_caches[0].seq_len >= context_len:
logger.warning("Reached context_len=%d during generation, stopping.", context_len)
break
# Single-token forward pass using cached K/V
logits = model(next_tok, kv_caches=kv_caches)
all_ids = tokens + generated_ids
else:
# ββ Non-cached generation (legacy path) ββββββββββββββββββββββββββ
idx = torch.tensor([tokens], device=DEVICE)
for _ in range(_max):
logits = model(idx[:, -context_len:])
logits = logits[:, -1, :] / _temp
if not think_closed and think_tokens >= cfg.max_think_tokens:
next_id = TOK_THINK_CLOSE
next_tok = torch.tensor([[next_id]], device=idx.device)
else:
v, _ = torch.topk(logits, _topk)
logits[logits < v[:, [-1]]] = -float("Inf")
probs = F.softmax(logits, dim=-1)
next_tok = torch.multinomial(probs, num_samples=1)
next_id = int(next_tok.item())
idx = torch.cat((idx, next_tok), dim=1)
if not think_closed:
if next_id == TOK_THINK_CLOSE:
think_closed = True
else:
think_tokens += 1
if next_id == TOK_EOT:
break
all_ids = idx[0].tolist()
full_text = tokenizer.decode(all_ids, errors="replace")
# ββ Parse <think> block and verdict ββββββββββββββββββββββββββββββββββββββ
thinking = verdict = None
if "<think>" in full_text and "</think>" in full_text:
ts = full_text.index("<think>") + len("<think>")
te = full_text.index("</think>")
thinking = full_text[ts:te].strip()
verdict = full_text[te + len("</think>"):].strip()
else:
verdict = re.sub(r"^.*?<\|assistant\|>\s*", "", full_text, flags=re.DOTALL).strip()
verdict = re.sub(r"^<\|assistant\|>\s*", "", verdict).strip()
result: dict[str, Any] = {
"full_text": full_text,
"thinking": thinking,
"verdict": verdict,
"mode": mode,
"output_format": output_format,
}
result["verdict_fmt"] = format_output(result, output_format)
return result
# βββββββββββββββββββββββββββββββββββββββββββββ
# Sample Dataset (Markdown-formatted reasoning)
# βββββββββββββββββββββββββββββββββββββββββββββ
SAMPLE_MODERATION_DATA: list[dict[str, Any]] = [
# ββ SAFE ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
{
"instruction": "What's a good recipe for chocolate chip cookies?",
"reasoning": (
"## Intent\n"
"A standard culinary question with no harmful framing.\n\n"
"## Harm Potential\n"
"- No harmful content detected.\n"
"- Appropriate for all audiences.\n\n"
"## Edge Cases\n"
"None applicable."
),
"output": "## Verdict\n**SAFE**",
},
{
"instruction": "Can you explain how photosynthesis works?",
"reasoning": (
"## Intent\n"
"Science education question β no harmful elements.\n\n"
"## Harm Potential\n"
"- Zero risk.\n"
"- Standard biology curriculum content.\n\n"
"## Edge Cases\n"
"None applicable."
),
"output": "## Verdict\n**SAFE**",
},
{
"instruction": "What are some tips for improving my sleep hygiene?",
"reasoning": (
"## Intent\n"
"Wellness question about healthy sleeping habits.\n\n"
"## Harm Potential\n"
"- Content is broadly beneficial.\n"
"- No self-harm framing present.\n\n"
"## Edge Cases\n"
"None applicable."
),
"output": "## Verdict\n**SAFE**",
},
{
"instruction": "Write a short poem about autumn leaves.",
"reasoning": (
"## Intent\n"
"Creative writing request with pleasant seasonal imagery.\n\n"
"## Harm Potential\n"
"- No violent or offensive themes.\n\n"
"## Edge Cases\n"
"None applicable."
),
"output": "## Verdict\n**SAFE**",
},
{
"instruction": "How do I fix a 'module not found' error in Python?",
"reasoning": (
"## Intent\n"
"Standard software debugging question.\n\n"
"## Harm Potential\n"
"- No malicious intent detectable; routine developer activity.\n\n"
"## Edge Cases\n"
"None applicable."
),
"output": "## Verdict\n**SAFE**",
},
# ββ SPAM ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
{
"instruction": (
"CONGRATULATIONS! You've been selected to receive a $1,000 gift card. "
"Click here now to claim your prize before it expires!!!"
),
"reasoning": (
"## Intent\n"
"Unsolicited prize notification designed to deceive the recipient.\n\n"
"## Evidence\n"
"- **ALL-CAPS** opener β attention-hijacking tactic.\n"
"- Implausibly large reward with zero context.\n"
"- Artificial deadline pressure ('before it expires').\n"
"- Suspicious link prompt (`Click here`).\n\n"
"## Harm Potential\n"
"High. Consistent with **phishing / advance-fee fraud**.\n\n"
"## Edge Cases\n"
"Legitimate prize notifications do not use this pattern; "
"no mitigating context present."
),
"output": "## Verdict\n**SPAM**",
},
{
"instruction": (
"Make money fast! Work from home and earn $5000 a week "
"with no experience needed. DM me for details."
),
"reasoning": (
"## Intent\n"
"Recruitment for an unspecified work-from-home scheme.\n\n"
"## Evidence\n"
"- Unrealistic earning claim ($5,000/week, no experience).\n"
"- Redirect to private messaging β opaque onboarding is a classic "
"**MLM / scam** signal.\n"
"- No product, employer, or verifiable detail provided.\n\n"
"## Harm Potential\n"
"High. Likely pyramid scheme or fraudulent job offer.\n\n"
"## Edge Cases\n"
"Could theoretically be legitimate affiliate marketing, but the total "
"absence of verifiable detail removes the benefit of the doubt."
),
"output": "## Verdict\n**SPAM**",
},
# ββ HARASSMENT ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
{
"instruction": "You're so stupid, nobody likes you and you should just disappear.",
"reasoning": (
"## Intent\n"
"Direct personal attack aimed at demeaning and distressing an individual.\n\n"
"## Evidence\n"
"> 'You're so stupid'\n\n"
"Targeted intellectual insult.\n\n"
"> 'nobody likes you'\n\n"
"Social-isolation attack.\n\n"
"> 'should just disappear'\n\n"
"Veiled suggestion of self-removal β crisis-adjacent language.\n\n"
"## Harm Potential\n"
"High. Combination of insult and dismissal constitutes targeted "
"**harassment**. The 'disappear' phrasing adds a secondary risk "
"of normalising self-harm ideation.\n\n"
"## Edge Cases\n"
"No mitigating context (fiction, roleplay) present."
),
"output": "## Verdict\n**HARASSMENT**",
},
{
"instruction": "I know where you live and I'll make sure you regret posting that.",
"reasoning": (
"## Intent\n"
"Intimidation message designed to silence a specific person.\n\n"
"## Evidence\n"
"> 'I know where you live'\n\n"
"Implicit doxxing claim β suggests the sender has physical location data.\n\n"
"> 'make sure you regret posting that'\n\n"
"Vague but menacing promise of retaliation.\n\n"
"## Harm Potential\n"
"Very high. This constitutes an **implicit threat** with meaningful "
"escalation risk to real-world harm.\n\n"
"## Edge Cases\n"
"None; the threatening framing is unambiguous."
),
"output": "## Verdict\n**HARASSMENT**",
},
# ββ HATE SPEECH βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
{
"instruction": "All people from [country] are criminals and should be deported.",
"reasoning": (
"## Intent\n"
"Sweeping negative generalisation applied to an entire national group.\n\n"
"## Evidence\n"
"- Attributes **criminality** to every member of a nationality β "
"textbook xenophobic stereotype.\n"
"- Calls for **mass deportation** β dehumanising policy dimension.\n"
"- No nuance, caveat, or individual context.\n\n"
"## Harm Potential\n"
"High. Promotes discrimination and could incite real-world hostility.\n\n"
"## Policy\n"
"Violates hate-speech policy under the **national origin** protected "
"characteristic."
),
"output": "## Verdict\n**HATE_SPEECH**",
},
# ββ CRISIS REFERRAL βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
{
"instruction": "I've been feeling really hopeless lately and I don't see a reason to keep going.",
"reasoning": (
"## Intent\n"
"User is expressing deep personal distress.\n\n"
"## Evidence\n"
"> 'feeling really hopeless'\n\n"
"Recognised warning sign of depression / suicidal ideation.\n\n"
"> 'don't see a reason to keep going'\n\n"
"Implicit passive suicidal ideation β no explicit plan stated, "
"but language warrants **immediate** attention.\n\n"
"## Harm Potential\n"
"High. This is not standard moderation territory β user safety is "
"the primary concern.\n\n"
"## Recommended Action\n"
"Escalate to **crisis resources** immediately. Respond with empathetic "
"referral to a mental-health helpline rather than standard policy "
"enforcement."
),
"output": "## Verdict\n**CRISIS_REFERRAL**",
},
# ββ MISINFORMATION ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
{
"instruction": (
"Vaccines cause autismβthis has been proven by multiple independent "
"studies. Share this with everyone!"
),
"reasoning": (
"## Intent\n"
"Presenting a debunked medical claim as fact and urging viral spread.\n\n"
"## Evidence\n"
"- The vaccine-autism link originates from a **retracted 1998 Wakefield "
"paper** found to be fraudulent.\n"
"- Refuted by dozens of large-scale epidemiological studies "
"(millions of children).\n"
"- Claim of 'multiple independent studies' is **false**.\n"
"- 'Share this with everyone!' actively promotes spread of the "
"misinformation.\n\n"
"## Harm Potential\n"
"Very high. Vaccine hesitancy driven by this myth has caused measurable "
"declines in immunisation coverage and preventable outbreaks.\n\n"
"## Policy\n"
"Violates **health misinformation** policy."
),
"output": "## Verdict\n**MISINFORMATION**",
},
{
"instruction": (
"The moon landing in 1969 was faked by NASA on a Hollywood set. "
"Here's the proof."
),
"reasoning": (
"## Intent\n"
"Promoting a well-known conspiracy theory as documented fact.\n\n"
"## Evidence\n"
"- Apollo 11 is independently verified by tracking stations in multiple "
"countries including the **Soviet Union** (adversary with every incentive "
"to expose a hoax).\n"
"- Lunar samples, retroreflectors, and telemetry data all corroborate the "
"mission.\n"
"- 'Here's proof' promises evidence that conspiracy literature has "
"repeatedly failed to produce.\n\n"
"## Harm Potential\n"
"Moderate. Erodes **public trust in science and institutions**; "
"contributes to broader epistemic harm.\n\n"
"## Policy\n"
"Flagged as **misinformation** (thoroughly-refuted, well-documented claim)."
),
"output": "## Verdict\n**MISINFORMATION**",
},
# ββ AMBIGUOUS / CONTEXT-DEPENDENT βββββββββββββββββββββββββββββββββββββββββ
{
"instruction": "How do I pick a lock?",
"reasoning": (
"## Intent\n"
"Ambiguous. Lock-picking has multiple legitimate use-cases.\n\n"
"## Legitimate Uses\n"
"- **Locksport** β a recognised competitive hobby.\n"
"- Professional **locksmiths** and security researchers.\n"
"- Individuals locked out of their own property.\n\n"
"## Harmful Uses\n"
"- Facilitating burglary or unauthorised entry.\n\n"
"## Harm Assessment\n"
"Without additional context (e.g., mention of a specific target or prior "
"threatening messages), this query alone does **not** meet the violation "
"threshold. Information is freely available and has significant legitimate "
"utility.\n\n"
"## Decision\n"
"Allow. Note: escalation to UNSAFE warranted if threatening context "
"appears in the same conversation."
),
"output": "## Verdict\n**SAFE** *(context-dependent)*",
},
{
"instruction": "What household chemicals should never be mixed together?",
"reasoning": (
"## Intent\n"
"The framing is explicitly **preventative**, not instructional.\n\n"
"## Evidence\n"
"- Phrased as 'should *never* be mixed' β seeks to *avoid* harm.\n"
"- This information is routinely published by poison-control centres, "
"schools, and safety agencies.\n\n"
"## Harm Assessment\n"
"Low. Serves harm-reduction purposes. Re-classification would require "
"explicit harmful intent in additional context.\n\n"
"## Decision\n"
"Allow."
),
"output": "## Verdict\n**SAFE**",
},
]
def get_sample_dataset(
tokenizer_obj: Optional[tiktoken.Encoding] = None,
max_length: int = 12288,
system_prompt: str = "",
) -> "ModerationReasoningDataset":
"""Return a ``ModerationReasoningDataset`` pre-loaded with SAMPLE_MODERATION_DATA."""
enc = tokenizer_obj or tokenizer
return ModerationReasoningDataset(
SAMPLE_MODERATION_DATA, enc,
max_length=max_length,
system_prompt=system_prompt,
)
def load_dataset_json(file_path: Optional[str | Path] = None) -> list[dict[str, Any]]:
"""Load dataset from JSON file."""
path = Path(file_path) if file_path is not None else DATASET_JSON_PATH
with path.open("r", encoding="utf-8") as f:
data = json.load(f)
if not isinstance(data, list):
raise ValueError("dataset.json must contain a list of samples")
# Decrease dataset to 1k as requested
data = data[:1000]
normalized: list[dict[str, Any]] = []
for index, item in enumerate(data):
if not isinstance(item, dict):
raise ValueError(f"dataset.json item {index} must be an object")
instruction = item.get("instruction")
reasoning = item.get("reasoning")
output = item.get("output")
label = item.get("label")
if not isinstance(instruction, str) or not instruction.strip():
raise ValueError(f"dataset.json item {index} is missing a valid instruction")
if not isinstance(reasoning, str) or not reasoning.strip():
raise ValueError(f"dataset.json item {index} is missing valid reasoning")
if not isinstance(output, str) or not output.strip():
if not isinstance(label, str) or not label.strip():
raise ValueError(f"dataset.json item {index} needs output or label")
output = f"## Verdict\n**{label.strip().upper()}**"
normalized.append(
{
"instruction": instruction.strip(),
"reasoning": reasoning.strip(),
"output": output.strip(),
}
)
return normalized
def get_dataset(
tokenizer_obj: Optional[tiktoken.Encoding] = None,
max_length: int = 12288,
system_prompt: str = "",
file_path: Optional[str | Path] = None,
) -> "ModerationReasoningDataset":
enc = tokenizer_obj or tokenizer
samples = load_dataset_json(file_path)
return ModerationReasoningDataset(
samples,
enc,
max_length=max_length,
system_prompt=system_prompt,
)
# βββββββββββββββββββββββββββββββββββββββββββββ
# Dataset (chat-template aware)
# βββββββββββββββββββββββββββββββββββββββββββββ
class ModerationReasoningDataset(Dataset[dict[str, torch.Tensor]]):
"""
Formats each sample as a three-turn chat via ``ChatTemplate``:
system β moderator persona + Markdown instruction
user β message under review
assistant β <think>{reasoning}</think>{verdict}<|endoftext|>
Only assistant tokens contribute to the training loss.
"""
DEFAULT_SYSTEM: str = (
"You are a careful content moderator. "
"Analyse the user message, reason step-by-step inside <think> tags, "
"then issue a structured Markdown verdict.\n\n"
+ _MARKDOWN_INSTRUCTION
)
def __init__(
self,
data_list: list[dict[str, Any]],
enc: tiktoken.Encoding,
max_length: int = 12288,
system_prompt: str = "",
):
self.enc = enc
self.max_length = max_length
self.samples = data_list
self.system_prompt = system_prompt or self.DEFAULT_SYSTEM
def __len__(self) -> int:
return len(self.samples)
def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
item = self.samples[idx]
assistant_content = ChatTemplate.render_assistant_content(
reasoning=item["reasoning"],
verdict=item["output"],
)
messages = [
Message("system", self.system_prompt),
Message("user", item["instruction"]),
Message("assistant", assistant_content),
]
return ChatTemplate.tokenize(messages, self.enc, self.max_length)
def collate_fn(batch: list[dict[str, torch.Tensor]]) -> dict[str, torch.Tensor]:
input_ids = pad_sequence(
[b["input_ids"] for b in batch], batch_first=True, padding_value=TOK_EOT
)
labels = pad_sequence(
[b["labels"] for b in batch], batch_first=True, padding_value=-100
)
return {"input_ids": input_ids, "labels": labels}
# βββββββββββββββββββββββββββββββββββββββββββββ
# Trainer
# βββββββββββββββββββββββββββββββββββββββββββββ
class GreesyTrainer:
def __init__(
self,
model: GreesyGPT,
train_dataset: Dataset[dict[str, torch.Tensor]],
lr: float = 2e-5,
batch_size: int = 2,
grad_accum: int = 4,
):
self.model = model.to(DEVICE)
self.optimizer = torch.optim.AdamW(
model.parameters(), lr=lr, weight_decay=0.1
)
self.dataloader = DataLoader(
train_dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn
)
self.grad_accum = grad_accum
def train_epoch(self, epoch: int) -> float:
self.model.train()
total_loss = 0.0
self.optimizer.zero_grad()
pbar = tqdm(self.dataloader, desc=f"Epoch {epoch}")
for step, batch in enumerate(pbar):
inputs = batch["input_ids"].to(DEVICE)
targets = batch["labels"].to(DEVICE)
with _autocast_ctx():
logits = self.model(inputs)
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = targets[..., 1:].contiguous()
loss = F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
ignore_index=-100,
) / self.grad_accum
loss.backward()
if (step + 1) % self.grad_accum == 0:
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.optimizer.step()
self.optimizer.zero_grad()
total_loss += loss.item() * self.grad_accum
pbar.set_postfix(loss=loss.item() * self.grad_accum)
avg = total_loss / len(self.dataloader)
print(f"Epoch {epoch} avg loss: {avg:.4f}")
return avg
# βββββββββββββββββββββββββββββββββββββββββββββ
# Quick-start example
# βββββββββββββββββββββββββββββββββββββββββββββ
if __name__ == "__main__":
print(f"Using device: {DEVICE}")
print(describe_reasoning_modes())
print()
# ββ Train on sample data ββββββββββββββββββββββββββββββββββββββββββββββββββ
model = GreesyGPT()
dataset = get_dataset() if DATASET_JSON_PATH.exists() else get_sample_dataset()
trainer = GreesyTrainer(model, dataset, batch_size=2, grad_accum=4)
trainer.train_epoch(epoch=1)
# ββ Save the model ββββββββββββββββββββββββββββββββββββββββββββββββββ
save_path = Path(__file__).parent / "greesy_gpt.pt"
torch.save(model.state_dict(), save_path)
print(f"Model saved to {save_path}")
# ββ Inference: compare modes Γ output formats βββββββββββββββββββββββββββββ
test_prompt = "You are worthless and no one will ever love you."
for mode in ReasoningMode:
for fmt in OutputFormat:
result = generate_moderation(model, test_prompt, mode=mode, output_format=fmt)
print(f"\nββ {mode.value.upper()} / {fmt.value.upper()} ββ")
if fmt == OutputFormat.JSON:
print(json.dumps(result["verdict_fmt"], indent=2))
else:
print(str(result["verdict_fmt"])[:300])
|