File size: 10,573 Bytes
c9dcc3b | 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 | from __future__ import annotations
import hashlib
import json
import random
from dataclasses import dataclass
from datetime import datetime, timezone
from pathlib import Path
from typing import Any
from datasets import Dataset, DatasetDict
from presidio_anonymizer.entities import OperatorConfig
from triage_llm.data.anonymize import PresidioAnonymizer
from triage_llm.schemas import DPORecord, MetadataSchema, SFTRecord
from triage_llm.utils import ensure_dir, read_jsonl, write_jsonl
@dataclass
class BuildDatasetsConfig:
input_dir: str
out_dir: str
seed: int = 42
split_ratios: tuple[float, float, float] = (0.9, 0.05, 0.05)
anonymize: bool = False
anonymize_lang_default: str = "fr"
anonymize_operator: str = "replace"
anonymize_new_value: str = "<REDACTED>"
export_hf: bool = True
clinical_eval_dir: str | None = None
def default_metadata_schema() -> MetadataSchema:
return MetadataSchema(
fields={
"id": "Identifiant unique",
"instruction/prompt": "Consigne ou prompt",
"input": "Contexte (optionnel)",
"output/chosen/rejected": "Réponse(s)",
"symptoms": "Liste de symptômes normalisés (optionnel)",
"history": "Antécédents (optionnel)",
"vitals": "Constantes (optionnel)",
"source": "Origine du dataset",
"lang": "Langue fr/en",
"confidence": "Niveau de confiance (optionnel)",
"pii_redacted": "PII supprimées (bool)",
}
)
def load_records_from_dir(input_dir: Path) -> tuple[list[SFTRecord], list[DPORecord]]:
sft: list[SFTRecord] = []
dpo: list[DPORecord] = []
for p in sorted(input_dir.glob("*.jsonl")):
rows = read_jsonl(p)
for row in rows:
if {"instruction", "output"}.issubset(row.keys()):
sft.append(SFTRecord.model_validate(row))
elif {"prompt", "chosen", "rejected"}.issubset(row.keys()):
dpo.append(DPORecord.model_validate(row))
return sft, dpo
def _sha256_file(path: Path) -> str:
h = hashlib.sha256()
with open(path, "rb") as f:
for chunk in iter(lambda: f.read(1024 * 1024), b""):
h.update(chunk)
return h.hexdigest()
def _now_utc_iso() -> str:
return datetime.now(timezone.utc).isoformat()
def _audit_write(path: Path, event: dict[str, Any]) -> None:
ensure_dir(path.parent)
with open(path, "a", encoding="utf-8") as f:
f.write(json.dumps(event, ensure_ascii=False) + "\n")
def _anonymize_sft_rows(
rows: list[dict[str, Any]],
lang_default: str,
operator: str,
new_value: str,
) -> tuple[list[dict[str, Any]], dict[str, Any]]:
op = OperatorConfig(operator, {"new_value": new_value})
engine_fr = PresidioAnonymizer(language="fr", operators={"DEFAULT": op})
engine_en = PresidioAnonymizer(language="en", operators={"DEFAULT": op})
n_entities_total = 0
out: list[dict[str, Any]] = []
for row in rows:
lang = (row.get("lang") or lang_default).lower()
engine = engine_fr if lang == "fr" else engine_en
for key in ["instruction", "input", "output"]:
if not row.get(key):
continue
try:
res = engine.anonymize(str(row[key]))
n_entities_total += len(res.entities)
row[key] = res.text
except Exception:
# Fallback: leave text unchanged but keep pipeline running.
row[key] = str(row[key])
row["pii_redacted"] = True
out.append(row)
stats = {"records": len(rows), "entities_detected": n_entities_total}
return out, stats
def _anonymize_dpo_rows(
rows: list[dict[str, Any]],
lang_default: str,
operator: str,
new_value: str,
) -> tuple[list[dict[str, Any]], dict[str, Any]]:
op = OperatorConfig(operator, {"new_value": new_value})
engine_fr = PresidioAnonymizer(language="fr", operators={"DEFAULT": op})
engine_en = PresidioAnonymizer(language="en", operators={"DEFAULT": op})
n_entities_total = 0
out: list[dict[str, Any]] = []
for row in rows:
lang = (row.get("lang") or lang_default).lower()
engine = engine_fr if lang == "fr" else engine_en
for key in ["prompt", "chosen", "rejected"]:
if not row.get(key):
continue
try:
res = engine.anonymize(str(row[key]))
n_entities_total += len(res.entities)
row[key] = res.text
except Exception:
row[key] = str(row[key])
row["pii_redacted"] = True
out.append(row)
stats = {"records": len(rows), "entities_detected": n_entities_total}
return out, stats
def split_rows(
rows: list[dict[str, Any]],
seed: int,
ratios: tuple[float, float, float] = (0.9, 0.05, 0.05),
):
assert abs(sum(ratios) - 1.0) < 1e-9
rng = random.Random(seed)
idx = list(range(len(rows)))
rng.shuffle(idx)
n = len(rows)
n_train = int(n * ratios[0])
n_val = int(n * ratios[1])
train = [rows[i] for i in idx[:n_train]]
val = [rows[i] for i in idx[n_train : n_train + n_val]]
test = [rows[i] for i in idx[n_train + n_val :]]
return train, val, test
def build_datasets(cfg: BuildDatasetsConfig) -> dict[str, Path]:
input_path = Path(cfg.input_dir)
out_path = ensure_dir(cfg.out_dir)
audit_path = out_path / "audit_log.jsonl"
_audit_write(
audit_path,
{
"ts": _now_utc_iso(),
"event": "build_start",
"input_dir": str(input_path),
"out_dir": str(out_path),
"seed": cfg.seed,
"split_ratios": cfg.split_ratios,
"anonymize": cfg.anonymize,
"export_hf": cfg.export_hf,
},
)
sft_records, dpo_records = load_records_from_dir(input_path)
sft_rows = [r.model_dump(mode="json") for r in sft_records]
dpo_rows = [r.model_dump(mode="json") for r in dpo_records]
_audit_write(
audit_path,
{
"ts": _now_utc_iso(),
"event": "loaded",
"sft_records": len(sft_rows),
"dpo_records": len(dpo_rows),
},
)
if cfg.anonymize:
sft_rows, sft_stats = _anonymize_sft_rows(
sft_rows,
lang_default=cfg.anonymize_lang_default,
operator=cfg.anonymize_operator,
new_value=cfg.anonymize_new_value,
)
dpo_rows, dpo_stats = _anonymize_dpo_rows(
dpo_rows,
lang_default=cfg.anonymize_lang_default,
operator=cfg.anonymize_operator,
new_value=cfg.anonymize_new_value,
)
_audit_write(
audit_path,
{
"ts": _now_utc_iso(),
"event": "anonymized",
"sft": sft_stats,
"dpo": dpo_stats,
"operator": cfg.anonymize_operator,
"new_value": cfg.anonymize_new_value,
},
)
sft_path = out_path / "sft.jsonl"
dpo_path = out_path / "dpo.jsonl"
write_jsonl(sft_path, sft_rows)
write_jsonl(dpo_path, dpo_rows)
schema = default_metadata_schema()
schema_path = out_path / "metadata_schema.json"
with open(schema_path, "w", encoding="utf-8") as f:
json.dump(schema.model_dump(mode="json"), f, ensure_ascii=False, indent=2)
splits_path = ensure_dir(out_path / "splits")
sft_train, sft_val, sft_test = split_rows(sft_rows, seed=cfg.seed, ratios=cfg.split_ratios)
dpo_train, dpo_val, dpo_test = split_rows(dpo_rows, seed=cfg.seed, ratios=cfg.split_ratios)
sft_train_path = splits_path / "sft_train.jsonl"
sft_val_path = splits_path / "sft_val.jsonl"
sft_test_path = splits_path / "sft_test.jsonl"
dpo_train_path = splits_path / "dpo_train.jsonl"
dpo_val_path = splits_path / "dpo_val.jsonl"
dpo_test_path = splits_path / "dpo_test.jsonl"
write_jsonl(sft_train_path, sft_train)
write_jsonl(sft_val_path, sft_val)
write_jsonl(sft_test_path, sft_test)
write_jsonl(dpo_train_path, dpo_train)
write_jsonl(dpo_val_path, dpo_val)
write_jsonl(dpo_test_path, dpo_test)
if cfg.export_hf:
sft_dd = DatasetDict(
{
"train": Dataset.from_list(sft_train),
"validation": Dataset.from_list(sft_val),
"test": Dataset.from_list(sft_test),
}
)
dpo_dd = DatasetDict(
{
"train": Dataset.from_list(dpo_train),
"validation": Dataset.from_list(dpo_val),
"test": Dataset.from_list(dpo_test),
}
)
hf_path = ensure_dir(out_path / "hf")
sft_hf_path = hf_path / "sft"
dpo_hf_path = hf_path / "dpo"
sft_dd.save_to_disk(str(sft_hf_path))
dpo_dd.save_to_disk(str(dpo_hf_path))
_audit_write(
audit_path,
{
"ts": _now_utc_iso(),
"event": "export_hf",
"sft_path": str(sft_hf_path),
"dpo_path": str(dpo_hf_path),
},
)
if cfg.clinical_eval_dir:
eval_in = Path(cfg.clinical_eval_dir)
eval_out = ensure_dir(out_path / "eval")
copied: list[str] = []
for p in sorted(eval_in.glob("*.jsonl")):
target = eval_out / p.name
target.write_text(p.read_text(encoding="utf-8"), encoding="utf-8")
copied.append(str(target))
_audit_write(
audit_path,
{"ts": _now_utc_iso(), "event": "eval_sets_copied", "files": copied},
)
_audit_write(
audit_path,
{
"ts": _now_utc_iso(),
"event": "build_end",
"outputs": {
"sft_jsonl": str(sft_path),
"dpo_jsonl": str(dpo_path),
"schema": str(schema_path),
"splits": str(splits_path),
"audit": str(audit_path),
},
"hashes": {
"sft_jsonl_sha256": _sha256_file(sft_path),
"dpo_jsonl_sha256": _sha256_file(dpo_path),
"schema_sha256": _sha256_file(schema_path),
},
},
)
return {
"sft": sft_path,
"dpo": dpo_path,
"schema": schema_path,
"splits": splits_path,
"audit": audit_path,
"hf": out_path / "hf" if cfg.export_hf else out_path,
}
|