Spaces:
Running
Running
File size: 15,850 Bytes
c31002d | 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 | """
Layer 2 — Anotação multi-annotator com famílias divergentes (Patch B).
Para cada par (q_bad, q_good) da Layer 1, três anotadores independentes
de arquiteturas distintas avaliam se a reformulação é genuinamente melhor.
Anotadores suportados:
- Claude Haiku (Anthropic) — ANTHROPIC_API_KEY [obrigatório]
- Llama 3.3 70B (Meta/Groq) — GROQ_API_KEY [recomendado]
- GPT-4o-mini (OpenAI) — OPENAI_API_KEY [opcional]
Saída: data/pairs/pairs_layer2.jsonl
Mantém apenas pares onde todos os anotadores ativos concordam
na direção (q_good > q_bad) E variância de magnitude < VAR_THRESHOLD.
Pares com divergência ficam em pairs_layer2_divergent.jsonl para auditoria.
"""
from __future__ import annotations
import json
import os
import re
import statistics
import time
from dataclasses import dataclass
from dataclasses import field
from pathlib import Path
from typing import Protocol
from dotenv import load_dotenv
from tqdm import tqdm
load_dotenv(override=True)
# Pares com variância de magnitude acima deste threshold vão para divergentes
VAR_THRESHOLD = 0.04 # stdev < 0.2 entre anotadores
MIN_ANNOTATORS = 2 # mínimo para considerar um par validado
# ---------------------------------------------------------------------------
# Estruturas de dados
# ---------------------------------------------------------------------------
@dataclass
class Annotation:
model: str
direction: bool # True = q_good é genuinamente melhor
magnitude: float # 0-1: quanto melhor
reasoning: str
error: str = ""
@dataclass
class AnnotatedPair:
# Campos originais da Layer 1
q_bad: str
q_good: str
context: str
domain: str
confidence: float
source: str
source_id: str
year: int
# Anotações da Layer 2
annotations: list[Annotation] = field(default_factory=list)
@property
def valid_annotations(self) -> list[Annotation]:
return [a for a in self.annotations if not a.error]
@property
def agreement_direction(self) -> bool:
va = self.valid_annotations
return len(va) >= MIN_ANNOTATORS and all(a.direction for a in va)
@property
def magnitude_variance(self) -> float:
va = self.valid_annotations
mags = [a.magnitude for a in va]
if len(mags) < 2:
return 0.0
return statistics.variance(mags)
@property
def magnitude_mean(self) -> float:
va = self.valid_annotations
if not va:
return 0.0
return statistics.mean(a.magnitude for a in va)
@property
def passes_layer2(self) -> bool:
return (
self.agreement_direction
and self.magnitude_variance <= VAR_THRESHOLD
and len(self.valid_annotations) >= MIN_ANNOTATORS
)
def to_dict(self) -> dict:
return {
"q_bad": self.q_bad,
"q_good": self.q_good,
"context": self.context,
"domain": self.domain,
"confidence": self.confidence,
"source": self.source,
"source_id": self.source_id,
"year": self.year,
"annotations": [
{
"model": a.model,
"direction": a.direction,
"magnitude": a.magnitude,
"reasoning": a.reasoning,
**({"error": a.error} if a.error else {}),
}
for a in self.annotations
],
"n_annotators": len(self.valid_annotations),
"agreement_direction": self.agreement_direction,
"magnitude_mean": round(self.magnitude_mean, 4),
"magnitude_variance": round(self.magnitude_variance, 4),
"passes_layer2": self.passes_layer2,
}
# ---------------------------------------------------------------------------
# Prompt compartilhado
# ---------------------------------------------------------------------------
_SYSTEM = """\
You are an expert in philosophy of science and research methodology.
Your task: evaluate whether a question reformulation is genuinely epistemically productive.
A reformulation is genuinely better (not just superficially different) when:
1. The new question has clearer methodology — there exist known tools, experiments, or formal frameworks to make progress on it.
2. Answering the new question makes meaningful progress toward answering the original question (inferential connection).
3. The new question is not merely a paraphrase, synonym substitution, or vocabulary upgrade of the original.
Respond ONLY with valid JSON. No markdown, no explanation outside the JSON.
{
"direction": <true if q_good is genuinely more epistemically tractable than q_bad, else false>,
"magnitude": <float 0.0-1.0 — how much better: 0=no improvement, 0.5=moderate, 1.0=transformative>,
"reasoning": "<one sentence explaining the key epistemic improvement, or why it fails>"
}
"""
_USER_TEMPLATE = """\
Context (from source text):
{context}
Original question (q_bad):
{q_bad}
Reformulated question (q_good):
{q_good}
Domain: {domain}
"""
def _parse_annotation(raw: str, model: str) -> Annotation:
raw = re.sub(r"```[a-z]*\n?", "", raw).strip()
try:
data = json.loads(raw)
return Annotation(
model=model,
direction=bool(data["direction"]),
magnitude=float(max(0.0, min(1.0, data["magnitude"]))),
reasoning=str(data.get("reasoning", "")),
)
except Exception as e:
return Annotation(model=model, direction=False, magnitude=0.0, reasoning="", error=str(e))
# ---------------------------------------------------------------------------
# Anotadores
# ---------------------------------------------------------------------------
class Annotator(Protocol):
name: str
def annotate(self, pair: dict) -> Annotation: ...
class ClaudeAnnotator:
name = "claude-haiku-4-5-20251001"
def __init__(self) -> None:
import anthropic
self._client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])
def annotate(self, pair: dict) -> Annotation:
prompt = _USER_TEMPLATE.format(**pair)
try:
msg = self._client.messages.create(
model=self.name,
max_tokens=256,
system=_SYSTEM,
messages=[{"role": "user", "content": prompt}],
)
return _parse_annotation(msg.content[0].text, self.name)
except Exception as e:
return Annotation(
model=self.name, direction=False, magnitude=0.0, reasoning="", error=str(e)
)
class GroqAnnotator:
name = "llama-3.3-70b-versatile"
def __init__(self) -> None:
from groq import Groq
self._client = Groq(api_key=os.environ["GROQ_API_KEY"])
def annotate(self, pair: dict) -> Annotation:
prompt = _USER_TEMPLATE.format(**pair)
try:
resp = self._client.chat.completions.create(
model=self.name,
max_tokens=256,
messages=[
{"role": "system", "content": _SYSTEM},
{"role": "user", "content": prompt},
],
temperature=0.1,
)
return _parse_annotation(resp.choices[0].message.content, self.name)
except Exception as e:
return Annotation(
model=self.name, direction=False, magnitude=0.0, reasoning="", error=str(e)
)
class OpenAIAnnotator:
name = "gpt-4o-mini"
def __init__(self) -> None:
from openai import OpenAI
self._client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
def annotate(self, pair: dict) -> Annotation:
prompt = _USER_TEMPLATE.format(**pair)
try:
resp = self._client.chat.completions.create(
model=self.name,
max_tokens=256,
messages=[
{"role": "system", "content": _SYSTEM},
{"role": "user", "content": prompt},
],
temperature=0.1,
response_format={"type": "json_object"},
)
return _parse_annotation(resp.choices[0].message.content, self.name)
except Exception as e:
return Annotation(
model=self.name, direction=False, magnitude=0.0, reasoning="", error=str(e)
)
class GeminiAnnotator:
name = "gemini-2.0-flash"
def __init__(self) -> None:
from google import genai
self._client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])
def annotate(self, pair: dict) -> Annotation:
from google.genai import types
prompt = _USER_TEMPLATE.format(**pair)
for attempt in range(4):
try:
resp = self._client.models.generate_content(
model=self.name,
contents=prompt,
config=types.GenerateContentConfig(
system_instruction=_SYSTEM,
temperature=0.1,
max_output_tokens=256,
),
)
return _parse_annotation(resp.text, self.name)
except Exception as e:
if "429" in str(e) and attempt < 3:
wait = 15 * (attempt + 1)
time.sleep(wait)
continue
return Annotation(
model=self.name, direction=False, magnitude=0.0, reasoning="", error=str(e)
)
return Annotation(
model=self.name,
direction=False,
magnitude=0.0,
reasoning="",
error="max retries exceeded",
)
def _build_annotators() -> list:
annotators = []
if os.getenv("ANTHROPIC_API_KEY"):
annotators.append(ClaudeAnnotator())
print(" [OK] Claude Haiku (Anthropic)")
else:
print(" [--] Claude — ANTHROPIC_API_KEY nao definida")
if os.getenv("GROQ_API_KEY"):
annotators.append(GroqAnnotator())
print(" [OK] Llama 3.3 70B (Groq)")
else:
print(" [--] Llama/Groq — GROQ_API_KEY nao definida (recomendado: console.groq.com)")
if os.getenv("GEMINI_API_KEY"):
annotators.append(GeminiAnnotator())
print(" [OK] Gemini 1.5 Flash (Google)")
else:
print(" [--] Gemini — GEMINI_API_KEY nao definida (gratuito: aistudio.google.com)")
if os.getenv("OPENAI_API_KEY"):
annotators.append(OpenAIAnnotator())
print(" [OK] GPT-4o-mini (OpenAI)")
else:
print(" [--] GPT — OPENAI_API_KEY nao definida (opcional)")
return annotators
# ---------------------------------------------------------------------------
# Pipeline principal
# ---------------------------------------------------------------------------
def annotate_pairs(
input_path: Path,
output_path: Path,
divergent_path: Path,
delay: float = 0.3,
) -> tuple[list[dict], list[dict]]:
pairs_raw = [
json.loads(l) for l in input_path.read_text(encoding="utf-8").splitlines() if l.strip()
]
print("\n=== Layer 2 — Multi-annotator ===")
print(f"Pares de entrada: {len(pairs_raw)}")
print("\nAnotadores ativos:")
annotators = _build_annotators()
if len(annotators) < MIN_ANNOTATORS:
raise RuntimeError(
f"Minimo de {MIN_ANNOTATORS} anotadores necessarios. "
f"Configure GROQ_API_KEY (gratuito em console.groq.com)."
)
# Retoma de onde parou
done_ids: set[str] = set()
agreed: list[dict] = []
divergent: list[dict] = []
for path, bucket in [(output_path, agreed), (divergent_path, divergent)]:
if path.exists():
for line in path.read_text(encoding="utf-8").splitlines():
if line.strip():
rec = json.loads(line)
done_ids.add(rec.get("source_id", ""))
bucket.append(rec)
if done_ids:
print(
f"\nRetomando: {len(done_ids)} pares ja anotados "
f"({len(agreed)} concordantes, {len(divergent)} divergentes)"
)
pending = [p for p in pairs_raw if p.get("source_id", "") not in done_ids]
print(f"Pares pendentes: {len(pending)}\n")
out_f = output_path.open("a", encoding="utf-8")
div_f = divergent_path.open("a", encoding="utf-8")
try:
for raw in tqdm(pending, desc="Anotando pares"):
ap = AnnotatedPair(
q_bad=raw["q_bad"],
q_good=raw["q_good"],
context=raw.get("context", ""),
domain=raw.get("domain", ""),
confidence=raw.get("confidence", 0.0),
source=raw.get("source", ""),
source_id=raw.get("source_id", ""),
year=raw.get("year", 0),
)
for ann in annotators:
ap.annotations.append(ann.annotate(raw))
time.sleep(delay)
record = ap.to_dict()
if ap.passes_layer2:
out_f.write(json.dumps(record, ensure_ascii=False) + "\n")
out_f.flush()
agreed.append(record)
else:
div_f.write(json.dumps(record, ensure_ascii=False) + "\n")
div_f.flush()
divergent.append(record)
finally:
out_f.close()
div_f.close()
return agreed, divergent
def print_summary(agreed: list[dict], divergent: list[dict]) -> None:
total = len(agreed) + len(divergent)
print(f"\n{'='*55}")
print(f" Total anotados: {total}")
print(f" Concordantes: {len(agreed)} ({100*len(agreed)//max(total,1)}%)")
print(f" Divergentes: {len(divergent)} ({100*len(divergent)//max(total,1)}%)")
if agreed:
mags = [p["magnitude_mean"] for p in agreed]
print(f" Magnitude media: {statistics.mean(mags):.3f}")
from collections import Counter
sources = Counter(p["source"] for p in agreed)
print("\n Por fonte (concordantes):")
for s, n in sources.most_common():
print(f" {s}: {n}")
# Motivos de divergência
if divergent:
no_dir = sum(1 for p in divergent if not p["agreement_direction"])
hi_var = sum(
1
for p in divergent
if p["agreement_direction"] and p["magnitude_variance"] > VAR_THRESHOLD
)
print("\n Motivo divergencia:")
print(f" Direcao discordante: {no_dir}")
print(f" Variancia alta: {hi_var}")
print(f"{'='*55}")
if __name__ == "__main__":
output_path = Path("data/pairs/pairs_layer2.jsonl")
divergent_path = Path("data/pairs/pairs_layer2_divergent.jsonl")
output_path.parent.mkdir(parents=True, exist_ok=True)
# Com Gemini (15 RPM free tier): delay de 5s entre pares garante < 12 RPM
has_gemini = bool(os.getenv("GEMINI_API_KEY"))
agreed, divergent = annotate_pairs(
input_path=Path("data/pairs/pairs_layer1.jsonl"),
output_path=output_path,
divergent_path=divergent_path,
delay=5.0 if has_gemini else 0.3,
)
print_summary(agreed, divergent)
print("\n=== Amostra concordantes ===")
for p in agreed[:3]:
models = [a["model"].split("-")[0] for a in p["annotations"]]
mags = [a["magnitude"] for a in p["annotations"]]
print(f"\n [{p['source']}] [{p['domain'][:40]}]")
print(f" BAD: {p['q_bad'][:80]}")
print(f" GOOD: {p['q_good'][:80]}")
print(f" Anotadores: {list(zip(models, [round(m,2) for m in mags]))}")
|