Spaces:
Running
Running
File size: 16,634 Bytes
a39d8ef | 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 | """
data_factory/generator.py
==========================
vLLM-based Natural Language question generator for H100.
This module uses a large LLM (Llama-3-70B or Qwen-72B) served via vLLM
to generate diverse, persona-based natural language paraphrases of the
canonical NL questions in our template library.
KEY DESIGN: The LLM generates ONLY natural language questions.
SQL is NEVER touched by the LLM.
This guarantees zero SQL errors in the final dataset.
Persona descriptions:
ceo - Direct, short, active voice. Business executive style.
chatty - Conversational, verbose, passive voice.
lazy_typist - Short, abbreviations, possible informal grammar.
non_techie - Plain English, avoids SQL/tech jargon, uses synonyms.
analyst - Technical, precise, jargon-heavy.
Usage (on H100 cluster):
python -m data_factory.generator --templates-per-chunk 20 --n-variants 10
"""
from __future__ import annotations
import json
import logging
import time
from typing import Iterator, Optional
logger = logging.getLogger(__name__)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# PERSONA SYSTEM PROMPTS
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
PERSONA_SYSTEM_PROMPTS: dict[str, str] = {
"ceo": (
"You are a busy C-level executive who communicates in short, punchy, "
"direct sentences. You use active voice, skip filler words, and get "
"straight to the point. You are asking a data analyst for information."
),
"chatty": (
"You are a friendly, conversational person who likes to be thorough "
"and explain things fully. You use passive voice sometimes, add context, "
"and ask questions in a relaxed, detailed way. You are not technical."
),
"lazy_typist": (
"You type quickly and informally. You use abbreviations (e.g. 'pls', "
"'lmk', 'asap'), lowercase, minimal punctuation, and sometimes omit "
"words. You get your meaning across without perfect grammar."
),
"non_techie": (
"You have no database or SQL knowledge. You use everyday English words "
"instead of technical terms. For example, you say 'customers' not 'rows', "
"'most expensive' not 'highest price', 'total money' not 'sum'. "
"You describe what you want to see, not how to get it."
),
"analyst": (
"You are a data scientist or BI analyst who is precise and technical. "
"You use terms like 'aggregate', 'partition', 'granularity', 'distinct', "
"'filter predicate', 'ranked by metric'. Your questions are precise and unambiguous."
),
}
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# PROMPT BUILDER
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def build_generation_prompt(
canonical_nl: str,
description: str,
persona: str,
schema_context: str,
n_variants: int = 5,
) -> list[dict[str, str]]:
"""
Build a chat-format prompt asking the LLM to rephrase the canonical NL
question in the style of the given persona.
Parameters
----------
canonical_nl : The base NL question from the template.
description : One-line SQL description (gives the LLM additional context).
persona : One of the 5 persona keys.
schema_context : The compact schema string for the domain.
n_variants : How many rephrased questions to generate.
Returns
-------
list[dict] Chat messages in [{"role": ..., "content": ...}] format.
"""
persona_desc = PERSONA_SYSTEM_PROMPTS[persona]
system = (
"You are a data labelling specialist. Your task is to rephrase a database "
"question in a specific communication style (persona). The rephrased questions "
"must preserve the EXACT same intent and required information as the original β "
"do not change what data is being asked for, only how it is expressed.\n\n"
f"PERSONA: {persona_desc}\n\n"
"OUTPUT FORMAT: Return ONLY a valid JSON array of strings. "
"No preamble, no markdown, no extra keys. Example: "
'["question 1", "question 2", "question 3"]'
)
user = (
f"DATABASE CONTEXT:\n{schema_context}\n\n"
f"WHAT THE QUERY DOES: {description}\n\n"
f"CANONICAL QUESTION: {canonical_nl}\n\n"
f"Generate {n_variants} different ways a person with the persona described "
f"above would ask this same question. The meaning must stay identical."
)
return [
{"role": "system", "content": system},
{"role": "user", "content": user},
]
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# RESPONSE PARSER
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def parse_llm_response(raw_text: str) -> list[str]:
"""
Extract a list of strings from the LLM's JSON response.
Handles common failures: markdown fences, trailing commas, extra text.
Returns an empty list if parsing fails completely.
"""
text = raw_text.strip()
# Strip markdown fences if present
if text.startswith("```"):
lines = text.split("\n")
text = "\n".join(l for l in lines if not l.strip().startswith("```")).strip()
# Find the JSON array boundaries
start = text.find("[")
end = text.rfind("]")
if start == -1 or end == -1 or end <= start:
logger.warning("LLM response missing JSON array brackets: %s", text[:100])
return []
json_str = text[start:end + 1]
# Fix trailing commas before ] (common LLM mistake)
json_str = json_str.rstrip()
json_str = json_str.replace(",]", "]").replace(", ]", "]")
try:
parsed = json.loads(json_str)
if not isinstance(parsed, list):
return []
# Filter to only non-empty strings
return [s.strip() for s in parsed if isinstance(s, str) and s.strip()]
except json.JSONDecodeError as exc:
logger.warning("JSON parse error: %s | text: %s", exc, json_str[:200])
return []
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# VLLM INTERFACE
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class VLLMGenerator:
"""
Wrapper around a running vLLM server for high-throughput NL generation.
Supports two modes:
online : Calls a running vLLM OpenAI-compatible API server.
offline : Uses vllm.LLM directly (loads model in-process, H100 recommended).
For H100 cluster usage, prefer 'offline' mode with tensor_parallel_size=4
to saturate all 4 H100s for maximum throughput.
"""
def __init__(
self,
model_name: str,
mode: str = "offline",
tensor_parallel_size: int = 4,
gpu_memory_utilization: float = 0.90,
max_model_len: int = 4096,
# Online mode only
api_base: str = "http://localhost:8000/v1",
api_key: str = "EMPTY",
) -> None:
self.model_name = model_name
self.mode = mode
self._llm = None
self._client = None
if mode == "offline":
self._init_offline(tensor_parallel_size, gpu_memory_utilization, max_model_len)
elif mode == "online":
self._init_online(api_base, api_key)
else:
raise ValueError(f"Unknown mode: {mode!r}. Use 'offline' or 'online'.")
def _init_offline(
self,
tensor_parallel_size: int,
gpu_memory_utilization: float,
max_model_len: int,
) -> None:
"""Load vLLM engine in-process (best for H100 cluster)."""
try:
from vllm import LLM, SamplingParams
self._LLM = LLM
self._SamplingParams = SamplingParams
except ImportError:
raise ImportError(
"vLLM not installed. Run: pip install vllm\n"
"For H100: pip install vllm --extra-index-url https://download.pytorch.org/whl/cu124"
)
logger.info("Loading model %s with %d GPUs (offline mode)...", self.model_name, tensor_parallel_size)
t0 = time.time()
self._llm = self._LLM(
model=self.model_name,
tensor_parallel_size=tensor_parallel_size,
gpu_memory_utilization=gpu_memory_utilization,
max_model_len=max_model_len,
dtype="bfloat16",
trust_remote_code=True,
)
logger.info("Model loaded in %.1f seconds.", time.time() - t0)
def _init_online(self, api_base: str, api_key: str) -> None:
"""Use OpenAI-compatible vLLM server (for distributed setups)."""
try:
from openai import OpenAI
self._client = OpenAI(base_url=api_base, api_key=api_key)
except ImportError:
raise ImportError("pip install openai")
logger.info("Connected to vLLM server at %s", api_base)
def generate_batch(
self,
prompts: list[list[dict[str, str]]],
temperature: float = 0.85,
max_new_tokens: int = 300,
) -> list[str]:
"""
Generate responses for a batch of chat prompts.
Parameters
----------
prompts : List of chat message lists (one per item in batch).
temperature : Sampling temperature. Higher = more diverse.
max_new_tokens : Max tokens per response.
Returns
-------
list[str] Raw text response per prompt (same length as input).
"""
if self.mode == "offline":
return self._generate_offline(prompts, temperature, max_new_tokens)
else:
return self._generate_online(prompts, temperature, max_new_tokens)
def _generate_offline(
self,
prompts: list[list[dict]],
temperature: float,
max_new_tokens: int,
) -> list[str]:
"""vLLM offline batched generation β maximises H100 throughput."""
from vllm import SamplingParams
sampling = SamplingParams(
temperature=temperature,
max_tokens=max_new_tokens,
stop=["</s>", "<|eot_id|>"], # Llama-3 stop tokens
)
# Convert chat messages to tokenised prompt strings using the model's template
tokenizer = self._llm.get_tokenizer()
formatted_prompts: list[str] = []
for msgs in prompts:
if hasattr(tokenizer, "apply_chat_template"):
text = tokenizer.apply_chat_template(
msgs, tokenize=False, add_generation_prompt=True
)
else:
# Fallback: simple concatenation
text = "\n".join(
f"<|{m['role']}|>\n{m['content']}" for m in msgs
)
formatted_prompts.append(text)
outputs = self._llm.generate(formatted_prompts, sampling)
return [o.outputs[0].text for o in outputs]
def _generate_online(
self,
prompts: list[list[dict]],
temperature: float,
max_new_tokens: int,
) -> list[str]:
"""Sequential generation via OpenAI-compatible API (fallback / debugging)."""
results = []
for msgs in prompts:
try:
resp = self._client.chat.completions.create(
model=self.model_name,
messages=msgs,
temperature=temperature,
max_tokens=max_new_tokens,
)
results.append(resp.choices[0].message.content or "")
except Exception as exc:
logger.warning("API call failed: %s", exc)
results.append("")
return results
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# HIGH-LEVEL GENERATION LOOP
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def generate_persona_variants_batch(
templates_subset: list[dict],
generator: VLLMGenerator,
personas: list[str],
n_variants_per_persona: int = 5,
batch_size: int = 64,
temperature: float = 0.85,
max_new_tokens: int = 300,
) -> Iterator[dict]:
"""
For each template Γ persona combination, generate `n_variants_per_persona`
NL question variants using the LLM.
Yields dicts:
{
"template_idx": int,
"persona": str,
"nl_variants": list[str], # successfully parsed NL questions
}
Parameters
----------
templates_subset : List of template dicts (from templates.py).
generator : VLLMGenerator instance.
personas : List of persona keys to use.
n_variants_per_persona : How many NL variants per (template, persona) pair.
batch_size : How many LLM calls to batch together.
temperature : Sampling temperature.
max_new_tokens : Max tokens for LLM response (should be ~300 for JSON array).
"""
from data_factory.schemas import SCHEMA_CONTEXT
# Build all (template_idx, persona) prompt pairs
all_jobs: list[tuple[int, str, list[dict]]] = []
for t_idx, template in enumerate(templates_subset):
schema_ctx = SCHEMA_CONTEXT[template["domain"]]
for persona in personas:
prompt = build_generation_prompt(
canonical_nl=template["base_nl"],
description=template["description"],
persona=persona,
schema_context=schema_ctx,
n_variants=n_variants_per_persona,
)
all_jobs.append((t_idx, persona, prompt))
total_jobs = len(all_jobs)
logger.info("Starting LLM generation: %d jobs (templates Γ personas).", total_jobs)
# Process in batches
for batch_start in range(0, total_jobs, batch_size):
batch = all_jobs[batch_start: batch_start + batch_size]
prompts = [job[2] for job in batch]
t0 = time.time()
raw_responses = generator.generate_batch(
prompts, temperature=temperature, max_new_tokens=max_new_tokens
)
elapsed = time.time() - t0
logger.info(
"Batch %d-%d completed in %.1fs (%.1f jobs/s).",
batch_start, batch_start + len(batch), elapsed, len(batch) / max(elapsed, 0.001)
)
for (t_idx, persona, _), raw in zip(batch, raw_responses):
nl_variants = parse_llm_response(raw)
if not nl_variants:
logger.debug(
"Empty parse for template_idx=%d persona=%s. raw=%s",
t_idx, persona, raw[:100]
)
# Fall back to the canonical NL rather than losing this entry
nl_variants = [templates_subset[t_idx]["base_nl"]]
yield {
"template_idx": t_idx,
"persona": persona,
"nl_variants": nl_variants,
}
|