File size: 10,414 Bytes
1b86964 | 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 | """Domain adapter interface for Chiron pipeline modules.
Active direct-prediction scripts load prompt modules explicitly, usually
``code/prompt/base/prediction_sft.py``.
"""
from __future__ import annotations
import json
import random
from abc import ABC, abstractmethod
from dataclasses import asdict, dataclass, field
from pathlib import Path
from typing import Any, Iterator
from utils.text_utils import (
format_generation_label,
normalize_prediction_label,
parse_batch_predictions,
parse_generation_predictions,
rouge_l_f1,
)
@dataclass
class Prediction:
"""A single prediction task for a sample."""
question: str
"""Full question text with options (with format instruction)."""
ground_truth: str
"""Correct answer label. New data should use bare numbers, e.g. '3'.
Legacy datasets may still contain 'Option 3'; training/eval utilities
normalize both forms to number-only labels.
"""
@dataclass
class MeasurementTarget:
"""A scalar psychometric target for representation shaping."""
construct: str
"""Stable construct name, for example ``big5.extraversion``."""
score: float
"""Normalized target score. Current joint-training code expects [-1, 1]."""
@dataclass
class Sample:
"""Domain-agnostic representation of one sample."""
observations: list[str]
predictions: list[Prediction]
data_source: str
split: str
within_person_holdout_predictions: list[Prediction] = field(default_factory=list)
extra_info: dict[str, Any] = field(default_factory=dict)
def to_dict(self) -> dict[str, Any]:
return asdict(self)
@classmethod
def from_dict(cls, d: dict[str, Any]) -> Sample:
predictions = [Prediction(**p) for p in d["predictions"]]
holdout_raw = d.get("within_person_holdout_predictions")
if holdout_raw is None:
holdout_raw = d.get("held_out_predictions", [])
holdout_predictions = [Prediction(**p) for p in holdout_raw]
obs = d["observations"]
if isinstance(obs, str):
obs = [s.strip() for s in obs.split("\n") if s.strip()]
return cls(
observations=obs,
predictions=predictions,
within_person_holdout_predictions=holdout_predictions,
data_source=d["data_source"],
split=d["split"],
extra_info=d.get("extra_info", {}),
)
def write_jsonl(samples: Iterator[Sample], path: Path) -> int:
"""Write samples to a JSONL file. Returns number of records written."""
path.parent.mkdir(parents=True, exist_ok=True)
count = 0
with path.open("w", encoding="utf-8") as f:
for sample in samples:
f.write(json.dumps(sample.to_dict(), ensure_ascii=False) + "\n")
count += 1
return count
def read_jsonl(path: Path) -> Iterator[Sample]:
"""Yield Sample instances from a JSONL file."""
with path.open("r", encoding="utf-8") as f:
for line in f:
line = line.strip()
if line:
yield Sample.from_dict(json.loads(line))
def _obs_body(observations: list[str]) -> str:
return "\n".join(f"- {o}" for o in observations)
def _questions_body(questions: list[str]) -> str:
"""Number questions starting at 1, separating with blank lines."""
return "\n\n".join(f"Q{i}.\n{q.strip()}" for i, q in enumerate(questions, start=1))
class DomainAdapter(ABC):
"""Lightweight interface so pipeline modules stay domain-agnostic."""
name: str = ""
"""Short domain name used by adapters and legacy prompt lookup."""
min_observations: int = 50
"""Default lower bound for sampled observations in direct prediction prompts."""
max_observations: int = 100
"""Default upper bound for sampled observations in direct prediction prompts."""
# ---- abstract: each domain must answer ----------------------------------
@abstractmethod
def data_source(self) -> str:
"""Dataset identifier, e.g. ``'indievalue_wvs'``."""
@abstractmethod
def get_sample_id(self, sample: Sample) -> str:
"""Return a unique string ID for *sample* (used for resume dedup)."""
@abstractmethod
def default_input_path(self) -> Path:
"""Default path (relative to repo root) for the train JSONL."""
# ---- task semantics -----------------------------------------------------
def task_type(self) -> str:
"""Prediction target type: ``classification`` or ``generation``."""
return "classification"
def metric_name(self) -> str:
"""Metric used by evolution-time scoring."""
return "exact_match"
def select_target_predictions(
self,
sample: Sample,
*,
seed: int,
sample_index: int,
max_questions: int,
min_questions: int = 1,
) -> list[Prediction]:
"""Select prediction questions used in one supervised direct-prediction row.
Domains should override this when the target-question sampling policy is
part of the dataset semantics. The base fallback shuffles predictions
deterministically and samples a question count uniformly from
``min_questions..min(max_questions, len(predictions))``. ``max_questions
<= 0`` means use all available predictions.
"""
preds = list(sample.predictions)
if not preds:
return []
rng = random.Random(seed ^ (sample_index * 1009))
rng.shuffle(preds)
limit = min(max_questions, len(preds)) if max_questions > 0 else len(preds)
floor = max(1, min(min_questions, limit))
n_target = rng.randint(floor, limit)
return preds[:n_target]
def select_measurement_targets(self, sample: Sample) -> list[MeasurementTarget]:
"""Return optional scalar targets for representation shaping.
Most direct-prediction domains do not expose continuous measurement
targets yet, so the default is empty. Domains can override this without
changing the direct-prediction answer format.
"""
del sample
return []
def select_observations(
self,
sample: Sample,
*,
seed: int,
sample_index: int,
) -> list[str]:
"""Select observation records used in one direct-prediction prompt.
Domains with very long record lists can override this hook. The default
mirrors the active direct-SFT policy: keep all observations when the
record is short; otherwise deterministically sample 50-100 records.
"""
observations = list(sample.observations)
if len(observations) <= self.min_observations:
return observations
rng = random.Random(seed ^ (sample_index * 9173) ^ 0x0B5E7A)
rng.shuffle(observations)
n_observations = rng.randint(
self.min_observations,
min(self.max_observations, len(observations)),
)
return observations[:n_observations]
def parse_prediction_output(
self, text: str, n: int, *, questions: list[str] | None = None,
) -> list[str]:
"""Parse model prediction output into one string per question."""
if self.task_type() == "generation":
return parse_generation_predictions(text, n)
return parse_batch_predictions(text, n)
def format_ground_truth(self, ground_truth: str) -> str:
"""Render a supervised answer label for training."""
if self.task_type() == "generation":
return format_generation_label(ground_truth)
return normalize_prediction_label(ground_truth)
def format_question_block(self, questions: list[str]) -> str:
"""Render selected questions in the user prompt."""
return _questions_body(questions)
def record_context(self) -> str:
"""One-sentence domain context shown immediately before the records tag."""
return "Below are records describing this person."
def format_observations_block(self, observations: list[str]) -> str:
"""Render observations in the user prompt."""
return _obs_body(observations)
def format_answer_block(self, predictions: list[Prediction]) -> str:
"""Render selected ground-truth labels in the assistant message."""
lines: list[str] = []
for i, pred in enumerate(predictions, start=1):
answer = self.format_ground_truth(pred.ground_truth)
if self.task_type() == "classification":
answer = f"Option {answer}"
lines.append(f"Q{i}: {answer}")
return "\n".join(lines)
def score_predictions(self, predictions: list[str], ground_truths: list[str]) -> dict[str, Any]:
"""Return a high-is-better score plus metric details."""
if self.task_type() == "generation":
scores = [rouge_l_f1(p, g) for p, g in zip(predictions, ground_truths)]
avg = sum(scores) / len(scores) if scores else 0.0
return {
"score": avg,
"rouge_l_f1": avg,
"per_example_rouge_l_f1": scores,
}
correct = sum(
normalize_prediction_label(p) == normalize_prediction_label(g)
for p, g in zip(predictions, ground_truths)
)
n = len(ground_truths)
return {"score": correct / n if n else 0.0, "n_correct": correct}
def prediction_feedback(
self,
questions: list[str],
predictions: list[str],
ground_truths: list[str],
) -> str:
"""Feedback block for mutation prompts."""
if self.task_type() == "generation":
blocks = []
for q, p, g in zip(questions, predictions, ground_truths):
score = rouge_l_f1(p, g)
blocks.append(
f"Q: {q}\n"
f" ROUGE-L F1: {score:.4f}\n"
f" Generated text: {p}\n"
f" Gold text: {g}"
)
return "\n\n".join(blocks) if blocks else "(none)"
wrong_items = [
f"Q: {q}\n Predicted: {p}\n Correct: {g}"
for q, p, g in zip(questions, predictions, ground_truths)
if normalize_prediction_label(p) != normalize_prediction_label(g)
]
return "\n\n".join(wrong_items) if wrong_items else "(none)"
|