File size: 31,335 Bytes
a1bf71f | 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 | #!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
chat_sprint_standalone.py
One-file pipeline: collect datasets -> reformat as You:/Bot: -> train tiny GPT (CUDA) -> sample & save model
Requirements (install once):
pip install torch datasets sentencepiece tqdm numpy
Run:
python chat_sprint_standalone.py
"""
import os, re, time, math, random, json
from pathlib import Path
from typing import List, Optional, Tuple
from itertools import islice
from contextlib import nullcontext
import numpy as np
from tqdm import tqdm
from datasets import load_dataset, get_dataset_config_names
import sentencepiece as spm
import torch
import torch.nn as nn
import torch.nn.functional as F
# --------------------------
# Global config (tweak here)
# --------------------------
SEED = 1337
random.seed(SEED); np.random.seed(SEED); torch.manual_seed(SEED)
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if DEVICE.type == "cuda":
torch.set_float32_matmul_precision("high")
torch.backends.cuda.matmul.allow_tf32 = True
SAVE_DIR = Path("./chat_sprint_artifacts")
SAVE_DIR.mkdir(parents=True, exist_ok=True)
# How many formatted You:/Bot: pairs to KEEP from each dataset (first run)
CAPS = {
"shakespeare": 15000,
"jokes": 20000,
"dadjokes": 8000,
"rsarcasm": 8000, # Thewillonline/reddit-sarcasm
"figlang": 12000,
"shower": 4000, # HuggingFaceGECLM/REDDIT_submissions split "Showerthoughts"
"personas": 2000,
"tweeteval": 4000,
"fourchan": 500, # dataset is tiny (~195)
"elonvtrump": 3000,
}
# Upper bound on rows to SCAN in each streaming dataset
SCAN = {
"jokes": 120_000,
"dadjokes": 60_000,
"rsarcasm": 120_000,
"figlang": 150_000,
"shower": 250_000,
"personas": 30_000,
"tweeteval": 60_000,
"fourchan": 2_000,
"elonvtrump": 60_000,
}
MAX_TOTAL_PAIRS = 60_000
MAX_LEN = 120
# Tokenizer
VOCAB_SIZE = 1500
TOKENIZER_PREFIX = SAVE_DIR / "spm_chat"
USER_SYMBOLS = ["You:", "Bot:", "[STYLE=Snark]", "[FORM=TWEET]", "[FORM=HEADLINE]", "[MOOD=Unhinged]", "[MOOD=Cheeky]"]
# Model size & train budget (~5 minutes on RTX 3090 with default)
block_size = 256
n_layer = 6
n_head = 6
n_embd = 384 # ~11.9M params
dropout = 0.0
MAX_SECONDS = 300 # hard cap
train_steps = 5000 # big number; time cap will stop near 5 min
log_interval = 100
eval_every = 400
batch_size = 24
accum_steps = 3
base_lr = 3e-3
min_lr = 5e-4
warmup_ratio = 0.06
# Sampling defaults
TEMP = 0.8
TOP_K = 60
TOP_P = 0.95
REP_PEN = 1.08
# --------------------------
# Helpers: cleaning & format
# --------------------------
URL_RE = re.compile(r"https?://\S+|www\.\S+", re.IGNORECASE)
MENT_RE = re.compile(r"@\w+")
WS_RE = re.compile(r"\s+")
QUOTE_RE = re.compile(r"^[\"'“”‘’]+|[\"'“”‘’]+$")
def clean_text(s: str) -> str:
s = s.strip()
s = URL_RE.sub("", s)
s = MENT_RE.sub("", s)
s = QUOTE_RE.sub("", s)
s = WS_RE.sub(" ", s)
return s.strip()
def shorten_to(s: str, n: int) -> str:
s = re.sub(r"\s+", " ", s.strip())
if len(s) <= n: return s
cut = max(s.rfind(". ", 0, n), s.rfind("! ", 0, n), s.rfind("? ", 0, n))
if cut != -1: return s[:cut+1].strip()
return s[:n].strip()
def keep_or_clip(s: str, min_len: int = 6, max_len: int = MAX_LEN) -> Optional[str]:
if not s: return None
s = re.sub(r"\s+", " ", s.strip())
if len(s) < min_len: return None
return shorten_to(s, max_len)
def turn(you: str, bot: str, tags: str = "") -> str:
lines = [f"You: {you}".rstrip()]
if tags: lines.append(tags)
lines.append(f"Bot: {bot}".rstrip())
lines.append("")
return "\n".join(lines)
def limited(ds, limit: int):
try:
return ds.take(limit)
except Exception:
return islice(ds, limit)
def get_first_nonempty(ex, keys) -> Optional[str]:
for k in keys:
v = ex.get(k)
if isinstance(v, str) and v.strip():
return v
return None
# --------------------------
# Collectors (console tqdm)
# --------------------------
def collect_shakespeare(pairs: List[str], overall: tqdm):
try:
ds = load_dataset(
"text",
data_files={"train": "https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt"},
split="train",
streaming=True,
)
kept = 0
pbar = tqdm(total=CAPS["shakespeare"], desc="[shakespeare]", unit="pair", leave=False, ncols=100)
for row in ds:
txt = keep_or_clip(clean_text(row["text"]))
if not txt: continue
pairs.append(turn("Continue in Shakespearean style.", txt))
kept += 1; pbar.update(1); overall.update(1)
if kept >= CAPS["shakespeare"]: break
pbar.close()
print(f"[ok] shakespeare kept={kept}")
except Exception as e:
print(f"[skip] shakespeare: {e}")
def collect_reddit_jokes(pairs: List[str], overall: tqdm):
for dsid in ["SocialGrep/one-million-reddit-jokes", "SocialGrep/reddit_jokes", "timc1/reddit_jokes"]:
try:
ds = load_dataset(dsid, split="train", streaming=True)
kept = 0
pbar = tqdm(total=CAPS["jokes"], desc="[jokes]", unit="pair", leave=False, ncols=100)
for ex in limited(ds, SCAN["jokes"]):
title = keep_or_clip(clean_text(str(ex.get("title") or "")))
body = keep_or_clip(clean_text(str(ex.get("selftext") or ex.get("body") or "")))
if body and title:
pairs.append(turn(title, body))
elif title:
pairs.append(turn("Tell me a short joke.", title))
else:
continue
kept += 1; pbar.update(1); overall.update(1)
if kept >= CAPS["jokes"]: break
pbar.close()
print(f"[ok] jokes {dsid} kept={kept}")
return
except Exception as e:
print(f"[try next] jokes {dsid}: {e}")
print("[skip] jokes: none worked")
def collect_dadjokes(pairs: List[str], overall: tqdm):
try:
ds = load_dataset("shuttie/reddit-dadjokes", split="train", streaming=True)
kept = 0
pbar = tqdm(total=CAPS["dadjokes"], desc="[dadjokes]", unit="pair", leave=False, ncols=100)
for ex in limited(ds, SCAN["dadjokes"]):
setup = keep_or_clip(clean_text(str(ex.get("setup") or ex.get("instruction") or ex.get("input") or "")))
punch = keep_or_clip(clean_text(str(ex.get("punchline") or ex.get("output") or "")))
if not (setup and punch): continue
pairs.append(turn(setup, punch))
kept += 1; pbar.update(1); overall.update(1)
if kept >= CAPS["dadjokes"]: break
pbar.close()
print(f"[ok] dadjokes kept={kept}")
except Exception as e:
print(f"[skip] dadjokes: {e}")
def collect_reddit_sarcasm(pairs: List[str], overall: tqdm):
"""Thewillonline/reddit-sarcasm — flexible parsing, scan+keep bars."""
try:
ds = load_dataset("Thewillonline/reddit-sarcasm", split="train", streaming=True)
keep_cap, scan_cap = CAPS["rsarcasm"], SCAN["rsarcasm"]
scanbar = tqdm(total=scan_cap, desc="[sarcasm scan]", unit="row", leave=False, ncols=100)
keepbar = tqdm(total=keep_cap, desc="[sarcasm kept]", unit="pair", leave=False, ncols=100)
PATS = [
re.compile(r"User\s*:\s*(.+?)\s*(?:Reddit\s*Comment|Comment|Reply)\s*:\s*(.+)", re.IGNORECASE | re.DOTALL),
re.compile(r"Post\s*:\s*(.+?)\s*(?:Top\s*Comment|Comment)\s*:\s*(.+)", re.IGNORECASE | re.DOTALL),
]
def parse(raw: str) -> Tuple[Optional[str], Optional[str]]:
raw = raw.replace("<|endoftext|>", "\n")
for pat in PATS:
m = pat.search(raw)
if m:
return m.group(1).strip(), m.group(2).strip()
lines = [ln.strip() for ln in raw.splitlines() if ln.strip()]
if len(lines) >= 2: return lines[0], lines[1]
if len(lines) == 1: return "Reply with sarcasm:", lines[0]
return None, None
kept = scanned = 0
for ex in limited(ds, scan_cap):
scanned += 1
you, bot = parse(str(ex.get("text") or ""))
you = keep_or_clip(you); bot = keep_or_clip(bot)
if you and bot:
pairs.append(turn(you, bot, "[STYLE=Snark]"))
kept += 1; keepbar.update(1); overall.update(1)
if kept >= keep_cap: break
scanbar.update(1)
if scanned % 2000 == 0:
keepbar.set_postfix(rate=f"{kept/max(1,scanned):.2%}")
scanbar.close(); keepbar.close()
print(f"[ok] reddit-sarcasm kept={kept} (scanned {scanned})")
except Exception as e:
print(f"[skip] reddit-sarcasm: {e}")
def collect_figlang(pairs: List[str], overall: tqdm):
for dsid in ["tasksource/figlang2020-sarcasm", "tasksource/figlang2020_sarcasm"]:
try:
ds = load_dataset(dsid, split="train", streaming=True)
kept = 0
pbar = tqdm(total=CAPS["figlang"], desc="[figlang]", unit="pair", leave=False, ncols=100)
for ex in limited(ds, SCAN["figlang"]):
ctx = ex.get("context")
if isinstance(ctx, list) and ctx:
context_str = " ".join(str(c) for c in ctx[-2:])
else:
context_str = str(ex.get("context") or ex.get("prompt") or "")
reply = str(ex.get("response") or ex.get("answer") or ex.get("text") or "")
context_str = keep_or_clip(clean_text(context_str))
reply = keep_or_clip(clean_text(reply))
if reply:
if context_str:
pairs.append(turn(context_str, reply, "[STYLE=Snark]"))
else:
pairs.append(turn("Reply with sarcasm:", reply, "[STYLE=Snark]"))
kept += 1; pbar.update(1); overall.update(1)
if kept >= CAPS["figlang"]: break
pbar.close()
print(f"[ok] figlang {dsid} kept={kept}")
return
except Exception as e:
print(f"[try next] figlang {dsid}: {e}")
print("[skip] figlang")
def collect_showerthoughts(pairs: List[str], overall: tqdm):
"""Use REEDIT_submissions split 'Showerthoughts' directly (no 'train')."""
try:
ds = load_dataset("HuggingFaceGECLM/REDDIT_submissions", split="Showerthoughts", streaming=True)
keep_cap, scan_cap = CAPS["shower"], SCAN["shower"]
scanbar = tqdm(total=scan_cap, desc="[shower scan]", unit="row", leave=False, ncols=100)
keepbar = tqdm(total=keep_cap, desc="[shower kept]", unit="pair", leave=False, ncols=100)
kept = scanned = 0
for ex in limited(ds, scan_cap):
scanned += 1
title = get_first_nonempty(ex, ["title", "selftext", "text"]) or ""
text = keep_or_clip(clean_text(title))
if text:
pairs.append(turn("Give me a shower thought.", text))
kept += 1; keepbar.update(1); overall.update(1)
if kept >= keep_cap: break
scanbar.update(1)
scanbar.close(); keepbar.close()
print(f"[ok] showerthoughts kept={kept} (scanned {scanned})")
except Exception as e:
print(f"[skip] showerthoughts: {e}")
def collect_personas(pairs: List[str], overall: tqdm):
"""Non-streaming is more reliable for this dataset."""
try:
ds = load_dataset("NapthaAI/twitter_personas")["train"]
keep_cap = CAPS["personas"]
pbar = tqdm(total=keep_cap, desc="[personas]", unit="pair", leave=False, ncols=100)
kept = 0
for ex in ds:
desc = get_first_nonempty(ex, ["description","persona","bio","text","content","full_text"])
if not isinstance(desc, str) and isinstance(ex.get("content"), dict):
desc = ex["content"].get("text")
desc = keep_or_clip(clean_text(str(desc or "")))
if not desc: continue
pairs.append(turn("Adopt this persona in one sentence.", desc, "[FORM=TWEET]"))
kept += 1; pbar.update(1); overall.update(1)
if kept >= keep_cap: break
pbar.close()
print(f"[ok] personas kept={kept}")
except Exception as e:
print(f"[skip] personas: {e}")
def collect_tweeteval(pairs: List[str], overall: tqdm):
"""Handle super_tweeteval (text_1/text_2, etc.) and fallback tweet_eval."""
def extract_pair(ex):
t = ex.get("text")
if isinstance(t, str) and t.strip():
return "React with a sharp one-liner.", t
for a,b in [("text_1","text_2"), ("sentence1","sentence2"),
("premise","hypothesis"), ("question","answer"),
("context","response"), ("tweet1","tweet2")]:
t1, t2 = ex.get(a), ex.get(b)
if isinstance(t1, str) and t1.strip() and isinstance(t2, str) and t2.strip():
return t1, t2
return None
def run_on(dsname, pick, is_super):
keep_cap, scan_cap = CAPS["tweeteval"], SCAN["tweeteval"]
pbar = tqdm(total=keep_cap, desc=f"[tweeteval:{pick}]", unit="pair", leave=False, ncols=100)
kept = 0
ds = load_dataset(dsname, pick, split="train", streaming=True)
for ex in limited(ds, scan_cap):
pair = extract_pair(ex) if is_super else ("React with a sharp one-liner.", ex.get("text")) if ex.get("text") else None
if not pair: continue
you, bot = pair
you = keep_or_clip(clean_text(str(you or ""))); bot = keep_or_clip(clean_text(str(bot or "")))
if not (you and bot): continue
tag = "[STYLE=Snark]" if you and you != "React with a sharp one-liner." else ""
pairs.append(turn(you, bot, tag))
kept += 1; pbar.update(1); overall.update(1)
if kept >= keep_cap: break
pbar.close()
return kept
kept_total = 0
try:
cfgs = get_dataset_config_names("cardiffnlp/super_tweeteval")
except Exception:
cfgs = []
prio = ["irony","sarcasm","humor","sentiment","emoji","emotion","stance","offensive","hate"]
ordered = [c for c in prio if c in cfgs] + [c for c in cfgs if c not in prio]
for pick in ordered:
kept_total += run_on("cardiffnlp/super_tweeteval", pick, True)
if kept_total >= CAPS["tweeteval"]:
print(f"[ok] tweeteval(super) kept={kept_total}"); return
if kept_total == 0:
try:
base_cfgs = get_dataset_config_names("cardiffnlp/tweet_eval")
except Exception:
base_cfgs = []
ordered_b = [c for c in prio if c in base_cfgs] + [c for c in base_cfgs if c not in prio]
for pick in ordered_b:
kept_total += run_on("cardiffnlp/tweet_eval", pick, False)
if kept_total >= CAPS["tweeteval"]:
print(f"[ok] tweeteval(base) kept={kept_total}"); return
print(f"[ok] tweeteval kept={kept_total}")
def collect_fourchan(pairs: List[str], overall: tqdm):
try:
ds = load_dataset("sbussiso/4chan-dataset", split="train", streaming=True)
keep_cap = min(CAPS["fourchan"], 195)
pbar = tqdm(total=keep_cap, desc="[4chan]", unit="pair", leave=False, ncols=100)
kept = 0
for ex in limited(ds, SCAN["fourchan"]):
prompt = keep_or_clip(clean_text(str(ex.get("prompt") or "")))
resp = keep_or_clip(clean_text(str(ex.get("response") or "")))
if prompt and resp:
pairs.append(turn(prompt, resp))
kept += 1; pbar.update(1); overall.update(1)
else:
txt = keep_or_clip(clean_text(str(ex.get("text") or ex.get("body") or ex.get("content") or "")))
if txt:
pairs.append(turn("Drop a spicy one-liner.", txt))
kept += 1; pbar.update(1); overall.update(1)
if kept >= keep_cap: break
pbar.close()
print(f"[ok] 4chan kept={kept}")
except Exception as e:
print(f"[skip] 4chan: {e}")
def collect_elon_trump(pairs: List[str], overall: tqdm):
try:
ds = load_dataset("MasaFoundation/Twitter_X_Elon_vs_Trump", split="train", streaming=True, revision="refs/convert/parquet")
keep_cap, scan_cap = CAPS["elonvtrump"], SCAN["elonvtrump"]
scanbar = tqdm(total=scan_cap, desc="[elon_vs_trump scan]", unit="row", leave=False, ncols=100)
keepbar = tqdm(total=keep_cap, desc="[elon_vs_trump kept]", unit="pair", leave=False, ncols=100)
kept = scanned = 0
for ex in limited(ds, scan_cap):
scanned += 1
txt = get_first_nonempty(ex, ["content","text","tweet","full_text"]) or ""
txt = keep_or_clip(clean_text(txt))
if txt:
pairs.append(turn("[FORM=TWEET] One sentence hot take:", txt, "[FORM=TWEET]"))
kept += 1; keepbar.update(1); overall.update(1)
if kept >= keep_cap: break
scanbar.update(1)
scanbar.close(); keepbar.close()
print(f"[ok] Elon_vs_Trump kept={kept} (scanned {scanned})")
except Exception as e:
print(f"[skip] Elon_vs_Trump: {e}")
# --------------------------
# Build corpus
# --------------------------
def build_corpus() -> Path:
pairs: List[str] = []
total_target = sum(CAPS.values())
print("[1/6] Collecting & reformatting datasets (streaming, capped)…")
overall = tqdm(total=total_target, desc="[all] collecting", unit="pair", ncols=100)
collectors = [
collect_shakespeare,
collect_reddit_jokes,
collect_dadjokes,
collect_reddit_sarcasm,
collect_figlang,
collect_showerthoughts,
collect_personas,
collect_tweeteval,
collect_fourchan,
collect_elon_trump,
]
for fn in collectors:
try:
fn(pairs, overall)
except Exception as e:
print(f"[collector error] {fn.__name__}: {e}")
overall.close()
print("[2/6] Deduplicating & clipping…")
seen = set(); deduped = []
for block in pairs:
try:
bot_line = [ln for ln in block.splitlines() if ln.startswith("Bot:")][0]
key = bot_line[4:].strip().lower()
except Exception:
key = block.strip().lower()
if key in seen: continue
seen.add(key); deduped.append(block)
random.shuffle(deduped)
if len(deduped) > MAX_TOTAL_PAIRS:
deduped = deduped[:MAX_TOTAL_PAIRS]
out_path = SAVE_DIR / "corpus.txt"
out_path.write_text("\n".join(deduped), encoding="utf-8")
print(f" wrote {len(deduped)} pairs → {out_path}")
return out_path
# --------------------------
# SentencePiece tokenizer
# --------------------------
def train_spm(corpus_path: Path) -> spm.SentencePieceProcessor:
print("[3/6] Training SentencePiece tokenizer…")
spm.SentencePieceTrainer.Train(
input=str(corpus_path),
model_prefix=str(TOKENIZER_PREFIX),
vocab_size=VOCAB_SIZE,
model_type="unigram",
character_coverage=1.0,
user_defined_symbols=USER_SYMBOLS,
bos_id=1, eos_id=2, unk_id=0, pad_id=-1
)
sp = spm.SentencePieceProcessor()
sp.load(f"{TOKENIZER_PREFIX}.model")
print(f" tokenizer saved at {TOKENIZER_PREFIX}.model")
return sp
# --------------------------
# Encode to token IDs
# --------------------------
def encode_corpus_to_ids(sp: spm.SentencePieceProcessor, corpus_path: Path):
print("[4/6] Encoding corpus to token IDs…")
text = corpus_path.read_text(encoding="utf-8")
blocks = [b for b in text.split("\n\n") if b.strip()]
ids = []
eos = sp.eos_id()
for b in blocks:
for line in b.splitlines():
if not line.strip(): continue
ids.extend(sp.encode(line, out_type=int)); ids.append(eos)
ids.append(eos)
ids = np.array(ids, dtype=np.int32)
n = len(ids); cut = int(n * 0.97)
train_ids = torch.tensor(ids[:cut], dtype=torch.long, device=DEVICE)
val_ids = torch.tensor(ids[cut:], dtype=torch.long, device=DEVICE)
print(f" tokens: train={train_ids.numel():,}, val={val_ids.numel():,}, vocab={sp.vocab_size()}")
return train_ids, val_ids, sp.vocab_size()
# --------------------------
# Tiny GPT model
# --------------------------
class CausalSelfAttention(nn.Module):
def __init__(self, n_embd, n_head, dropout=0.0, block_size=256):
super().__init__()
assert n_embd % n_head == 0
self.n_head = n_head
self.head_dim = n_embd // n_head
self.qkv = nn.Linear(n_embd, 3*n_embd, bias=False)
self.proj = nn.Linear(n_embd, n_embd, bias=False)
self.attn_drop = nn.Dropout(dropout)
self.resid_drop = nn.Dropout(dropout)
self.register_buffer("mask", torch.tril(torch.ones(block_size, block_size)).view(1,1,block_size,block_size))
def forward(self, x):
B,T,C = x.shape
qkv = self.qkv(x); q,k,v = qkv.chunk(3, dim=-1)
q = q.view(B,T,self.n_head,self.head_dim).transpose(1,2)
k = k.view(B,T,self.n_head,self.head_dim).transpose(1,2)
v = v.view(B,T,self.n_head,self.head_dim).transpose(1,2)
att = (q @ k.transpose(-2,-1)) / math.sqrt(self.head_dim)
att = att.masked_fill(self.mask[:,:,:T,:T] == 0, float('-inf'))
att = torch.softmax(att, dim=-1)
att = self.attn_drop(att)
y = att @ v
y = y.transpose(1,2).contiguous().view(B,T,C)
y = self.resid_drop(self.proj(y))
return y
class Block(nn.Module):
def __init__(self, n_embd, n_head, dropout=0.0, block_size=256):
super().__init__()
self.ln1 = nn.LayerNorm(n_embd)
self.attn = CausalSelfAttention(n_embd, n_head, dropout, block_size)
self.ln2 = nn.LayerNorm(n_embd)
self.mlp = nn.Sequential(
nn.Linear(n_embd, 4*n_embd),
nn.GELU(),
nn.Linear(4*n_embd, n_embd),
nn.Dropout(dropout),
)
def forward(self, x):
x = x + self.attn(self.ln1(x))
x = x + self.mlp(self.ln2(x))
return x
class TinyGPT(nn.Module):
def __init__(self, vocab_size, n_layer, n_head, n_embd, block_size, dropout=0.0):
super().__init__()
self.block_size = block_size
self.tok_emb = nn.Embedding(vocab_size, n_embd)
self.pos_emb = nn.Embedding(block_size, n_embd)
self.blocks = nn.ModuleList([Block(n_embd, n_head, dropout, block_size) for _ in range(n_layer)])
self.ln_f = nn.LayerNorm(n_embd)
self.head = nn.Linear(n_embd, vocab_size, bias=False)
self.apply(self._init)
def _init(self, m):
if isinstance(m, (nn.Linear, nn.Embedding)): nn.init.normal_(m.weight, mean=0.0, std=0.02)
if isinstance(m, nn.Linear) and m.bias is not None: nn.init.zeros_(m.bias)
if isinstance(m, nn.LayerNorm): nn.init.ones_(m.weight); nn.init.zeros_(m.bias)
def forward(self, idx, targets=None):
B,T = idx.shape; assert T <= self.block_size
pos = torch.arange(0, T, device=idx.device)
x = self.tok_emb(idx) + self.pos_emb(pos)[None,:,:]
for blk in self.blocks: x = blk(x)
x = self.ln_f(x)
logits = self.head(x)
loss = None
if targets is not None:
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
return logits, loss
@torch.no_grad()
def generate(self, idx, max_new_tokens=200, temperature=0.8, top_k=50, top_p=0.95, repetition_penalty=1.0):
self.eval()
for _ in range(max_new_tokens):
idx_cond = idx[:, -self.block_size:]
logits, _ = self.forward(idx_cond)
logits = logits[:, -1, :]
if repetition_penalty != 1.0:
uniq, _ = torch.unique(idx_cond[0], return_counts=True)
logits[:, uniq] /= repetition_penalty
logits = logits / max(1e-8, temperature)
if top_k is not None:
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
cutoff = v[:, -1].unsqueeze(-1)
logits = torch.where(logits < cutoff, torch.full_like(logits, -1e9), logits)
if top_p is not None:
sorted_logits, sorted_idx = torch.sort(logits, descending=True)
probs = torch.softmax(sorted_logits, dim=-1)
cdf = torch.cumsum(probs, dim=-1)
mask = cdf > top_p; mask[:, 0] = False
sorted_logits[mask] = -1e9
logits = torch.zeros_like(logits).scatter(1, sorted_idx, sorted_logits)
probs = torch.softmax(logits, dim=-1)
next_id = torch.multinomial(probs, num_samples=1)
idx = torch.cat([idx, next_id], dim=1)
return idx
# --------------------------
# Data loader for token IDs
# --------------------------
def get_batch(split_ids: torch.Tensor, B: int, T: int):
ix = torch.randint(0, split_ids.numel() - T - 1, (B,), device=split_ids.device)
x = torch.stack([split_ids[i:i+T] for i in ix])
y = torch.stack([split_ids[i+1:i+T+1] for i in ix])
return x, y
# --------------------------
# Train loop
# --------------------------
def train_model(vocab_size, train_ids, val_ids):
print("[5/6] Training tiny GPT on", DEVICE.type.upper(), "…")
model = TinyGPT(vocab_size, n_layer, n_head, n_embd, block_size, dropout).to(DEVICE)
params_m = sum(p.numel() for p in model.parameters())/1e6
print(f" params: {params_m:.2f}M")
optimizer = torch.optim.AdamW(model.parameters(), lr=base_lr, betas=(0.9, 0.95), weight_decay=0.0)
use_amp = DEVICE.type == "cuda"
scaler = torch.amp.GradScaler("cuda", enabled=use_amp)
autocast = (lambda: torch.amp.autocast("cuda", dtype=torch.float16)) if use_amp else nullcontext
start = time.time()
best_val = float("inf")
def get_lr(step):
warmup = max(1, int(train_steps * warmup_ratio))
if step < warmup: return base_lr * (step+1)/warmup
progress = (step - warmup) / max(1, train_steps - warmup)
return min_lr + 0.5*(base_lr - min_lr)*(1 + math.cos(math.pi * min(1.0, progress)))
@torch.no_grad()
def eval_loss(iters=80):
model.eval(); losses=[]
for _ in range(iters):
xb, yb = get_batch(val_ids, min(batch_size, 32), block_size)
with autocast():
_, loss = model(xb, yb)
losses.append(loss.item())
model.train()
return float(sum(losses)/len(losses))
model.train(); step = 0
pbar = tqdm(total=train_steps, ncols=100, desc="[train]")
while step < train_steps and (time.time()-start) < MAX_SECONDS:
lr = get_lr(step)
for pg in optimizer.param_groups: pg["lr"] = lr
optimizer.zero_grad(set_to_none=True)
total_loss = 0.0
for _ in range(accum_steps):
xb, yb = get_batch(train_ids, batch_size, block_size)
with autocast():
_, loss = model(xb, yb)
if use_amp:
scaler.scale(loss / accum_steps).backward()
else:
(loss / accum_steps).backward()
total_loss += loss.item()
if use_amp:
scaler.unscale_(optimizer)
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
if use_amp:
scaler.step(optimizer); scaler.update()
else:
optimizer.step()
step += 1; pbar.update(1)
if step % log_interval == 0 or step == 1:
pbar.set_postfix(train=f"{total_loss/accum_steps:.3f}", lr=f"{lr:.2e}")
if step % eval_every == 0:
vl = eval_loss()
best_val = min(best_val, vl)
print(f" eval loss {vl:.3f} | best {best_val:.3f}")
pbar.close()
elapsed = time.time() - start
print(f" done in {elapsed:.1f}s | best val {best_val:.3f}")
# Save model + config
ckpt_path = SAVE_DIR / "tinygpt.pt"
torch.save(model.state_dict(), ckpt_path)
(SAVE_DIR / "model_config.json").write_text(json.dumps({
"vocab_size": int(vocab_size),
"n_layer": n_layer, "n_head": n_head, "n_embd": n_embd,
"block_size": block_size, "dropout": dropout
}, indent=2))
print(f"[saved] weights → {ckpt_path}")
return model
# --------------------------
# Sampling helper
# --------------------------
def sample_chat(sp: spm.SentencePieceProcessor, model: TinyGPT, prompt: str, max_new_tokens=200):
prefix = f"You: {prompt}\nBot:"
ids = sp.encode(prefix, out_type=int)
x = torch.tensor([ids], dtype=torch.long, device=DEVICE)
with torch.no_grad():
y = model.generate(x, max_new_tokens=max_new_tokens, temperature=TEMP, top_k=TOP_K, top_p=TOP_P, repetition_penalty=REP_PEN)
return sp.decode(y[0].tolist())
# --------------------------
# Main
# --------------------------
def main():
# Build or reuse corpus/tokenizer/ids
corpus_path = SAVE_DIR / "corpus.txt"
spm_model = SAVE_DIR / "spm_chat.model"
if not corpus_path.exists():
corpus_path = build_corpus()
else:
print(f"[cache] using {corpus_path}")
sp = spm.SentencePieceProcessor()
if not spm_model.exists():
sp = train_spm(corpus_path)
else:
sp.load(str(spm_model))
print(f"[cache] using {spm_model}")
enc_train = SAVE_DIR / "train_ids.pt"
enc_val = SAVE_DIR / "val_ids.pt"
vocab_txt = SAVE_DIR / "vocab_size.txt"
if enc_train.exists() and enc_val.exists() and vocab_txt.exists():
train_ids = torch.load(enc_train, map_location=DEVICE)
val_ids = torch.load(enc_val, map_location=DEVICE)
vocab_size = int(vocab_txt.read_text())
print(f"[cache] loaded ids: train={train_ids.numel():,}, val={val_ids.numel():,}, vocab={vocab_size}")
else:
train_ids, val_ids, vocab_size = encode_corpus_to_ids(sp, corpus_path)
torch.save(train_ids, enc_train); torch.save(val_ids, enc_val)
vocab_txt.write_text(str(vocab_size))
print("[cache] saved encoded ids")
model = train_model(vocab_size, train_ids, val_ids)
print("\n[6/6] Samples:\n")
prompts = [
"Give me a spicy take on AI.",
"Roast my messy desk.",
"Explain recursion like you're annoyed.",
"Write a satirical headline about coffee.",
"Give me a shower thought about umbrellas.",
"Tell me a one-liner about deadlines.",
"Stay in Shakespeare mode and flatter me.",
"Reply sarcastically to: I love meetings.",
]
out_path = SAVE_DIR / "samples.txt"
with out_path.open("w", encoding="utf-8") as f:
for p in prompts:
txt = sample_chat(sp, model, p, max_new_tokens=200)
print("----\n" + txt)
f.write("----\n" + txt + "\n")
print(f"\n[saved] samples → {out_path}")
if __name__ == "__main__":
main()
|