File size: 20,647 Bytes
3af75d1 | 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 | import os
import json
import math
import asyncio
import random
from typing import Dict, Any, List, Tuple, Optional
from openai import AsyncOpenAI, BadRequestError
from tqdm import tqdm
from itertools import count
from prompts_resumo import CATEGORIES_SYSTEM_PROMPTS
# ---------------------------------------------------------------------
# Config vLLM OpenAI-compatible
# ---------------------------------------------------------------------
VLLM_BASE_URL = os.environ.get("VLLM_BASE_URL", "http://10.100.0.111:8022/v1")
VLLM_API_KEY = os.environ.get("VLLM_API_KEY", "no-key-needed")
CATEGORIES_SYS_PROMPTS = CATEGORIES_SYSTEM_PROMPTS
# Decoding (globais)
GEN_TEMPERATURE = float(os.environ.get("GEN_TEMPERATURE", "0.7"))
GEN_TOP_P = float(os.environ.get("GEN_TOP_P", "1.0"))
GEN_MAX_NEW_TOK = int(os.environ.get("GEN_MAX_NEW_TOK", "8192"))
STOP_STRINGS = ["<|im_end|>", "<|end_of_text|>"]
STOP_TOKEN_IDS = None
# Decoding diferenciada (resposta vs pergunta)
RESP_TEMPERATURE = float(os.environ.get("RESP_TEMPERATURE", str(GEN_TEMPERATURE if GEN_TEMPERATURE else 0.3)))
RESP_TOP_P = float(os.environ.get("RESP_TOP_P", str(min(GEN_TOP_P, 0.9))))
RESP_MAX_TOKENS = int(os.environ.get("RESP_MAX_TOKENS", "8192"))
Q_TEMPERATURE = float(os.environ.get("Q_TEMPERATURE", "0.7"))
Q_TOP_P = float(os.environ.get("Q_TOP_P", "0.95"))
Q_MAX_TOKENS = int(os.environ.get("Q_MAX_TOKENS", "8192"))
# ---------------------------------------------------------------------
# Controle de execução (apenas single-turn)
# ---------------------------------------------------------------------
NUM_ROWS = int(os.environ.get("NUM_ROWS", "300000"))
BATCH_SIZE = int(os.environ.get("BATCH_SIZE", "16"))
OUTPUT_FILE = os.environ.get("OUTPUT_FILE", "cemig_summary.jsonl")
INCLUDE_SYSTEM = True # (não será salvo no output)
LOGITS_PROCESSORS: List[str] = []
# Concorrência
MAX_ASYNC_TOTAL = int(os.environ.get("MAX_ASYNC", "32"))
CHAT_SHARE = float(os.environ.get("CHAT_SHARE", "0.8")) # 80% chat / 20% qgen
MAX_ASYNC_CHAT = max(1, int(MAX_ASYNC_TOTAL * CHAT_SHARE))
MAX_ASYNC_QGEN = max(1, MAX_ASYNC_TOTAL - MAX_ASYNC_CHAT)
QUEUE_MAXSIZE = int(os.environ.get("QUEUE_MAXSIZE", str(MAX_ASYNC_CHAT * 4)))
# ---------------------------------------------------------------------
# Dataset Wikipedia em streaming (HuggingFaceFW/clean-wikipedia, 'pt')
# ---------------------------------------------------------------------
WIKI_DATASET_ID = os.environ.get("WIKI_DATASET_ID", "cemig-ceia/energy_dataset_v1")
WIKI_SUBSET = os.environ.get("WIKI_SUBSET", "default")
WIKI_TEXT_FIELD = os.environ.get("WIKI_TEXT_FIELD", "text")
WIKI_MAX_CHARS = int(os.environ.get("WIKI_MAX_CHARS", "1500000"))
WIKI_MIN_CHARS = int(os.environ.get("WIKI_MIN_CHARS", "100"))
# ---------------------------------------------------------------------
# Estilos de PERGUNTA (mantidos)
# ---------------------------------------------------------------------
HARD_PROMPT_PAIR_SHARE = float(os.environ.get("HARD_PROMPT_PAIR_SHARE", "0.05"))
USER_PROMPTS = [
"Me ajude a responder essa questão, me dê uma resposta que explica o passo a passo do raciocinio.",
"Quero ajuda com essa questão mas sou ruim nessa materia, me explique ela mas também falando brevemente dos fundamentos dela.",
"Responda a questão de forma breve"
]
QUESTION_STYLE_PROMPTS = {
"summary": (
"Siga em pt-BR. Gere UMA pergunta solicitando resumo ou síntese de algo já mencionado "
"na conversa. A saída deve ser uma PERGUNTA. Evite sim/não e perguntas vagas."
"Máx. 1-2 sentenças; termine com '?'."
),
"rewrite_simplify": (
"Siga em pt-BR. Gere UMA pergunta solicitando reescrever o texto de forma que simplifique o jargão e vocabulário jurídico"
"A saída deve ser uma PERGUNTA. Evite sim/não e perguntas vagas."
"Máx. 1-2 sentenças; termine com '?'."
),
"rewrite_detailed": (
"Siga em pt-BR. Gere UMA pergunta solicitando reescrever de forma mais detalhada "
"algo já mencionado na conversa. A saída deve ser uma PERGUNTA. Evite sim/não e perguntas vagas."
"Máx. 1-2 sentença; termine com '?'."
),
"rewrite_focus": (
"Siga em pt-BR. Gere UMA pergunta solicitando reescrever algo já discutido, focando em um aspecto ou parte específica da resposta anterior. "
"A saída deve ser uma PERGUNTA. Evite sim/não e perguntas vagas. Máx. 1 sentença; termine com '?'."
),
"rewrite_formal": (
"Siga em pt-BR. Gere UMA pergunta solicitando reescrever algo já discutido na conversa de forma mais formal. "
"A saída deve ser uma PERGUNTA. Evite sim/não e perguntas vagas."
),
}
_DEFAULT_STYLE_PROBS = {
"summary": 0.100,
"rewrite_simplify": 0.00,
"rewrite_detailed": 0.00,
"rewrite_focus": 0.00,
"rewrite_formal": 0.00,
}
QUESTION_SAMPLES_PER_ROW = int(os.environ.get("QUESTION_SAMPLES_PER_ROW", "2"))
try:
QUESTION_STYLE_PROBS = json.loads(os.environ.get("QUESTION_STYLE_PROBS", "")) or _DEFAULT_STYLE_PROBS
except Exception:
QUESTION_STYLE_PROBS = _DEFAULT_STYLE_PROBS
# ---------------------------------------------------------------------
# Client
# ---------------------------------------------------------------------
client = AsyncOpenAI(base_url=VLLM_BASE_URL, api_key=VLLM_API_KEY)
# ---------------------------------------------------------------------
# Utilidades (sem regex)
# ---------------------------------------------------------------------
def _normalize_spaces(s: str) -> str:
return " ".join((s or "").split())
def pick_two_prompts(pool): # 2 hardcoded
pool = [p.strip() for p in pool if p and p.strip()]
if not pool:
return None
if len(pool) == 1:
return pool[0], pool[0]
return tuple(random.sample(pool, 2))
def _truncate_context(txt: str) -> str:
if not isinstance(txt, str):
return ""
t = _normalize_spaces(txt)
if len(t) <= WIKI_MAX_CHARS:
return t
cut = t.rfind(".", 0, WIKI_MAX_CHARS)
if cut == -1 or cut < WIKI_MIN_CHARS:
return t[:WIKI_MAX_CHARS]
return t[:cut+1]
def sample_system_prompt_key(prompts_with_probs: Dict[str, Tuple[str, float, str]]) -> Tuple[str, str]:
keys = list(prompts_with_probs.keys())
probs = [prompts_with_probs[k][1] for k in keys]
s = sum(probs)
probs = [1.0/len(keys)]*len(keys) if s <= 0 else [p/s for p in probs]
selected_key = random.choices(keys, weights=probs, k=1)[0]
context_type = prompts_with_probs[selected_key][2]
return selected_key, context_type
async def get_model_id() -> str:
models = await client.models.list()
if not models.data:
raise RuntimeError("Nenhum modelo disponível no endpoint vLLM.")
return models.data[0].id
async def chat_call(
messages: List[Dict[str, str]],
model_id: str,
*,
extra_body_override: Optional[dict] = None,
) -> str:
final_messages = messages
temperature = RESP_TEMPERATURE
top_p = RESP_TOP_P
max_tokens = RESP_MAX_TOKENS
extra_body = {
"chat_template_kwargs": {"enable_thinking": False},
"stop": STOP_STRINGS,
"stop_token_ids": STOP_TOKEN_IDS,
"logits_processors": LOGITS_PROCESSORS,
}
if extra_body_override:
for k, v in extra_body_override.items():
extra_body[k] = v
resp = await client.chat.completions.create(
model=model_id,
messages=final_messages,
temperature=temperature,
top_p=top_p,
max_tokens=max_tokens,
extra_body=extra_body,
)
return resp.choices[0].message.content or ""
async def safe_chat_call(*args, **kwargs) -> Optional[str]:
try:
return await chat_call(*args, **kwargs)
except BadRequestError:
try:
eb = kwargs.pop("extra_body_override", {}) or {}
eb2 = {"chat_template_kwargs": {"enable_thinking": True}}
eb2.update(eb)
return await chat_call(*args, extra_body_override=eb2, **kwargs)
except Exception:
pass
try:
eb = kwargs.pop("extra_body_override", {}) or {}
eb2 = dict(eb)
eb2.pop("chat_template_kwargs", None)
return await chat_call(*args, extra_body_override=eb2, **kwargs)
except Exception:
return None
except Exception:
return None
# ---------------------------------------------------------------------
# Geração de PERGUNTA — otimizada (n=2/JSON/fallback paralelo)
# ---------------------------------------------------------------------
def _sample_question_style(k=2) -> List[str]:
keys = list(QUESTION_STYLE_PROMPTS.keys())
weights = [QUESTION_STYLE_PROBS.get(k, 0.0) for k in keys]
s = sum(weights)
weights = [1.0/len(keys)]*len(keys) if s <= 0 else [w/s for w in weights]
return random.choices(keys, weights=weights, k=k)
async def generate_user_questions_pair(model_id: str, context_text: str, styles: List[str]) -> List[str]:
s0, s1 = styles[0], styles[1]
if s0 == s1:
try:
sys = (
"Você é um GERADOR de PERGUNTAS. Use APENAS o CONTEXTO. "
"Saída: UMA pergunta em PT-BR, sem comentários."
)
msg = [{"role": "system", "content": sys + f"\n\nINSTRUÇÃO DE ESTILO: {QUESTION_STYLE_PROMPTS[s0]}\n\nCONTEXTO:\n{context_text}"}]
resp = await client.chat.completions.create(
model=model_id,
messages=msg,
temperature=Q_TEMPERATURE,
top_p=Q_TOP_P,
max_tokens=Q_MAX_TOKENS,
n=2,
)
outs = [(c.message.content or "").strip() for c in resp.choices]
if len(outs) == 2 and all(outs):
return outs
except Exception:
pass
try:
sys = (
"Você é um GERADOR de PERGUNTAS. Use APENAS o CONTEXTO. "
"Produza JSON exatamente no formato: {\"q1\": \"...\", \"q2\": \"...\"}."
)
style_block = (
f"q1_style: {QUESTION_STYLE_PROMPTS[s0]}\n"
f"q2_style: {QUESTION_STYLE_PROMPTS[s1]}"
)
msg = [{"role": "system", "content": f"{sys}\n\n{style_block}\n\nCONTEXTO:\n{context_text}"}]
resp = await client.chat.completions.create(
model=model_id,
messages=msg,
temperature=Q_TEMPERATURE,
top_p=Q_TOP_P,
max_tokens=Q_MAX_TOKENS * 2,
)
txt = (resp.choices[0].message.content or "").strip()
q1 = q2 = None
try:
obj = json.loads(txt)
q1 = (obj.get("q1") or "").strip()
q2 = (obj.get("q2") or "").strip()
except Exception:
parts = [p.strip() for p in txt.split("\n") if p.strip()]
if len(parts) >= 2:
q1, q2 = parts[0], parts[1]
if q1 and q2:
return [q1, q2]
except Exception:
pass
async def _once(sty: str):
sys = (
"Você é um GERADOR de PERGUNTAS. Use APENAS o CONTEXTO. "
"Saída: apenas UMA pergunta em PT-BR, sem comentários."
)
msg = [{"role": "system", "content": sys + f"\n\nINSTRUÇÃO DE ESTILO: {QUESTION_STYLE_PROMPTS[sty]}\n\nCONTEXTO:\n{context_text}"}]
r = await client.chat.completions.create(
model=model_id,
messages=msg,
temperature=Q_TEMPERATURE,
top_p=Q_TOP_P,
max_tokens=Q_MAX_TOKENS,
)
return (r.choices[0].message.content or "").strip()
q1, q2 = await asyncio.gather(_once(s0), _once(s1))
return [q1, q2]
# ---------------------------------------------------------------------
# Conversa (single-turn: apenas user/assistant no output)
# ---------------------------------------------------------------------
async def generate_one_conversation(
model_id: str,
system_prompt_key: str,
system_prompt_text: str,
context_text: str,
initial_user: str,
question_style: str,
context_type: str,
) -> Dict[str, Any]:
conversation: List[Dict[str, str]] = []
base_msgs: List[Dict[str, str]] = []
if INCLUDE_SYSTEM and system_prompt_text:
base_msgs.append({"role": "system", "content": system_prompt_text})
if context_text and initial_user:
user_content = f"{context_text}\n\n{initial_user}"
elif context_text:
user_content = context_text
else:
user_content = initial_user or ""
conversation.append({"role": "user", "content": user_content})
first_answer = await safe_chat_call(base_msgs + [{"role": "user", "content": user_content}], model_id)
conversation.append({"role": "assistant", "content": first_answer})
return {"conversation": conversation}
# ---------------------------------------------------------------------
# Streaming do dataset + helpers
# ---------------------------------------------------------------------
def _wiki_stream_iter():
from datasets import load_dataset
return load_dataset(WIKI_DATASET_ID, WIKI_SUBSET, split="train", streaming=True)
def _extract_context(record: Dict[str, Any]) -> Optional[Dict[str, Any]]:
txt = record.get(WIKI_TEXT_FIELD, "")
if not isinstance(txt, str):
return None
context_text = _truncate_context(txt)
if len(context_text) < WIKI_MIN_CHARS:
return None
title = record.get("title", "")
return {"context_text": context_text, "title": title}
async def _execute_with_concurrency(coros: List[asyncio.Future], metas: List[dict], limit: int):
sem = asyncio.Semaphore(limit)
async def _runner(idx: int, coro):
async with sem:
res = await coro
return idx, res
tasks = [asyncio.create_task(_runner(i, c)) for i, c in enumerate(coros)]
try:
for done in asyncio.as_completed(tasks):
i, res = await done
yield res, metas[i]
finally:
for t in tasks:
if not t.done():
t.cancel()
def _read_resume_state(path: str) -> Tuple[int, int]:
if not os.path.exists(path) or os.path.getsize(path) == 0:
return 0, -1
next_seq_id = 0
last_ctx_id = -1
with open(path, "r", encoding="utf-8") as fin:
for line in fin:
try:
obj = json.loads(line)
except Exception:
continue
if "seq_id" in obj:
next_seq_id = max(next_seq_id, int(obj["seq_id"]) + 1)
if "context_id" in obj:
last_ctx_id = max(last_ctx_id, int(obj["context_id"]))
return next_seq_id, last_ctx_id
async def _skip_accepted_contexts(ds_iter, n_to_skip: int) -> int:
skipped = 0
if n_to_skip <= 0:
return 0
for rec in ds_iter:
if _extract_context(rec):
skipped += 1
if skipped >= n_to_skip:
break
return skipped
# ---------------------------------------------------------------------
# Runner com pipeline produtor→fila→consumidores (80/20 chat)
# ---------------------------------------------------------------------
async def run():
os.makedirs(os.path.dirname(OUTPUT_FILE) or ".", exist_ok=True)
model_id = await get_model_id()
# === RESUME ===
next_seq_id, last_ctx_id = _read_resume_state(OUTPUT_FILE)
seq_id = next_seq_id
ds_iter = _wiki_stream_iter()
_ = await _skip_accepted_contexts(ds_iter, last_ctx_id + 1)
start_ctx = (last_ctx_id + 1) if last_ctx_id >= 0 else 0
global_id = count(start_ctx)
total_remaining = max(0, NUM_ROWS - seq_id)
total_batches = math.ceil(total_remaining / BATCH_SIZE) if BATCH_SIZE > 0 else 0
# Abra em append
with open(OUTPUT_FILE, "a", encoding="utf-8") as fout:
pbar = tqdm(total=total_batches, desc="Gerando conversas (Magpie + Wikipedia streaming)")
write_lock = asyncio.Lock()
while seq_id < NUM_ROWS:
# Limita contextos por batch (2 conversas/ctx)
convs_restantes = NUM_ROWS - seq_id
max_contexts_este_batch = min(BATCH_SIZE, max(1, math.ceil(convs_restantes / 2)))
# Coleta contextos aceitos
batch_contexts: List[dict] = []
for rec in ds_iter:
context_dict = _extract_context(rec)
if context_dict:
ctx_id = next(global_id)
batch_contexts.append({**context_dict, "ctx_sample_id": ctx_id})
if len(batch_contexts) >= max_contexts_este_batch:
break
if not batch_contexts:
break # fim do stream
# Estilo da RESPOSTA por batch
key, context_type = sample_system_prompt_key(CATEGORIES_SYS_PROMPTS)
sys_prompt_text = CATEGORIES_SYS_PROMPTS[key][0]
# ===== Pipeline =====
conv_queue: asyncio.Queue = asyncio.Queue(maxsize=QUEUE_MAXSIZE)
# PRODUTORES (QGEN)
qgen_sem = asyncio.Semaphore(MAX_ASYNC_QGEN)
async def qgen_runner(cinfo: dict):
async with qgen_sem:
ctx_text = cinfo["context_text"]
ctx_id = cinfo["ctx_sample_id"]
use_hard_pair = (random.random() < HARD_PROMPT_PAIR_SHARE) and USER_PROMPTS
if use_hard_pair:
chosen = pick_two_prompts(USER_PROMPTS)
if chosen:
q1, q2 = chosen
styles = ["hardcoded", "hardcoded"]
else: #fallback caso algo dê errado
styles = _sample_question_style(k=2)
q1, q2 = await generate_user_questions_pair(model_id, ctx_text, styles)
else:
styles = _sample_question_style(k=2)
q1, q2 = await generate_user_questions_pair(model_id, ctx_text, styles)
await conv_queue.put((ctx_id, ctx_text, q1, styles[0]))
await conv_queue.put((ctx_id, ctx_text, q2, styles[1]))
producers = [asyncio.create_task(qgen_runner(cinfo)) for cinfo in batch_contexts]
# CONSUMIDORES (CHAT)
async def chat_worker(worker_id: int):
nonlocal seq_id
while True:
item = await conv_queue.get()
if item is None:
conv_queue.task_done()
break
ctx_id, ctx_text, question, q_style = item
try:
async with write_lock:
if seq_id >= NUM_ROWS:
return
r = await generate_one_conversation(
model_id=model_id,
system_prompt_key=key,
system_prompt_text=sys_prompt_text,
context_text=ctx_text,
initial_user=question,
question_style=q_style,
context_type=context_type,
)
async with write_lock:
if seq_id < NUM_ROWS:
record = {
"seq_id": seq_id,
"conversation": r["conversation"],
"question_style": q_style,
"context_id": ctx_id,
}
try:
fout.write(json.dumps(record, ensure_ascii=False) + "\n")
except Exception:
pass
seq_id += 1
except Exception:
pass
finally:
conv_queue.task_done()
consumers = [asyncio.create_task(chat_worker(i)) for i in range(MAX_ASYNC_CHAT)]
# Aguarda produtores concluírem e envia sentinelas
await asyncio.gather(*producers)
for _ in range(MAX_ASYNC_CHAT):
await conv_queue.put(None)
# Aguarda fila esvaziar e consumidores finalizarem
await conv_queue.join()
await asyncio.gather(*consumers, return_exceptions=True)
pbar.update(1)
pbar.close()
if __name__ == "__main__":
asyncio.run(run())
|