File size: 18,402 Bytes
f440f03 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 | """Human-in-the-loop Space helpers for staging training artifacts."""
from __future__ import annotations
import json
import os
from datetime import UTC, datetime
from pathlib import Path
from typing import Any
from uuid import uuid4
from pydantic import BaseModel, ConfigDict, Field, field_validator, model_validator
from maris_core.data.quality import (
DatasetQualityGateConfig,
apply_quality_gate_to_records,
build_dataset_quality_report,
)
from maris_core.training.preferences import (
PreferenceExample,
build_blind_side_by_side_artifact,
build_human_eval_summary,
summarize_preference_dataset,
)
from maris_core.training.space_ui import SAFE_OUTPUT_SEGMENT_RE
from maris_core.utils.env import validate_hf_model, validate_hf_repo_id
HUMAN_TRAINING_STAGE_DIRNAME = "human-training-staging"
HUMAN_TRAINING_REPO_PREFIX = "human-training"
HUMAN_TRAINING_MIN_TEXT_CHARS = 24
HUMAN_TRAINING_BLIND_REVIEW_SEED = 7
def _timestamp() -> str:
return datetime.now(UTC).replace(microsecond=0).isoformat()
def _new_run_id() -> str:
timestamp = datetime.now(UTC).strftime("%Y%m%dT%H%M%SZ")
return f"{timestamp}-{uuid4().hex[:8]}"
def _validate_output_subdir(value: str) -> str:
normalized = value.strip().strip("/")
if not normalized:
raise ValueError("Output apakšdirektorija nedrīkst būt tukša.")
parts = Path(normalized).parts
if ".." in parts or not SAFE_OUTPUT_SEGMENT_RE.fullmatch(normalized):
raise ValueError("Output apakšdirektorijā drīkst būt tikai droši ceļa segmenti.")
return normalized
def _normalize_lines(values: list[str]) -> list[str]:
normalized: list[str] = []
seen: set[str] = set()
for value in values:
text = str(value or "").strip()
if not text:
continue
signature = text.casefold()
if signature in seen:
continue
seen.add(signature)
normalized.append(text)
return normalized
class HumanTrainingConversationExample(BaseModel):
model_config = ConfigDict(str_strip_whitespace=True)
user: str = Field(min_length=1)
assistant: str = Field(min_length=1)
context: str = ""
class HumanTrainingPreferencePair(BaseModel):
model_config = ConfigDict(str_strip_whitespace=True)
prompt: str = Field(min_length=1)
chosen: str = Field(min_length=1)
rejected: str = Field(min_length=1)
context: str = ""
confidence: float | None = Field(default=None, ge=0.0, le=1.0)
class HumanTrainingEvalExample(BaseModel):
model_config = ConfigDict(str_strip_whitespace=True)
prompt: str = Field(min_length=1)
completion: str = Field(min_length=1)
context: str = ""
class HumanTrainingRequest(BaseModel):
"""Request payload for building staged human-training artifacts."""
model_config = ConfigDict(str_strip_whitespace=True)
dataset_repo: str
model_repo: str = ""
hub_model_id: str = ""
model_preset: str = "balanced"
model_name: str = ""
num_epochs: int = Field(default=3, ge=1, le=100)
all_branches: bool = False
push_to_hub: bool = True
output_subdir: str = "maris-human-training"
continue_from_latest_artifact: bool = True
continue_model_path: str = ""
profile_facts: list[str] = Field(default_factory=list)
profile_preferences: list[str] = Field(default_factory=list)
response_instructions: list[str] = Field(default_factory=list)
conversation_examples: list[HumanTrainingConversationExample] = Field(default_factory=list)
preference_pairs: list[HumanTrainingPreferencePair] = Field(default_factory=list)
eval_examples: list[HumanTrainingEvalExample] = Field(default_factory=list)
@field_validator("dataset_repo")
@classmethod
def validate_dataset_repo(cls, value: str) -> str:
try:
return validate_hf_repo_id(value, "dataset_repo", label="dataset repo")
except RuntimeError as exc:
raise ValueError(str(exc)) from exc
@field_validator("model_repo")
@classmethod
def validate_model_repo(cls, value: str) -> str:
normalized = value.strip()
if not normalized:
return ""
try:
return validate_hf_model(normalized, "model_repo")
except RuntimeError as exc:
raise ValueError(str(exc)) from exc
@field_validator("hub_model_id")
@classmethod
def validate_hub_model_id(cls, value: str) -> str:
normalized = value.strip()
if not normalized:
return ""
try:
return validate_hf_model(normalized, "hub_model_id")
except RuntimeError as exc:
raise ValueError(str(exc)) from exc
@field_validator("model_name")
@classmethod
def validate_model_name(cls, value: str) -> str:
normalized = value.strip()
if not normalized:
return ""
try:
return validate_hf_model(normalized, "model_name")
except RuntimeError as exc:
raise ValueError(str(exc)) from exc
@field_validator("output_subdir")
@classmethod
def validate_output_subdir(cls, value: str) -> str:
return _validate_output_subdir(value)
@field_validator("continue_model_path")
@classmethod
def validate_continue_model_path(cls, value: str) -> str:
normalized = value.strip()
if not normalized:
return ""
return _validate_output_subdir(normalized)
@field_validator("profile_facts", "profile_preferences", "response_instructions")
@classmethod
def dedupe_lines(cls, value: list[str]) -> list[str]:
return _normalize_lines(value)
@model_validator(mode="after")
def validate_has_training_signal(self) -> HumanTrainingRequest:
resolved_model_repo = self.hub_model_id or self.model_repo
if not resolved_model_repo:
raise ValueError("Jānorāda hub_model_id vai model_repo.")
self.hub_model_id = resolved_model_repo
self.model_repo = resolved_model_repo
if not any(
(
self.profile_facts,
self.profile_preferences,
self.response_instructions,
self.conversation_examples,
self.preference_pairs,
self.eval_examples,
)
):
raise ValueError("Human training pieprasījumā jābūt vismaz vienam ievades blokam.")
if not self.model_name and not self.model_preset:
self.model_preset = "balanced"
return self
class HumanTrainingExecuteRequest(BaseModel):
model_config = ConfigDict(str_strip_whitespace=True)
run_id: str = Field(min_length=6)
publish_artifacts: bool = True
start_training: bool = False
@model_validator(mode="after")
def validate_execution_flags(self) -> HumanTrainingExecuteRequest:
if self.start_training and not self.publish_artifacts:
raise ValueError(
"Lai sāktu treniņu, artefakti vispirms jāpublicē dataset repozitorijā."
)
return self
class HumanTrainingLaunchSpec(BaseModel):
model_config = ConfigDict(str_strip_whitespace=True)
dataset_repo: str
model_repo: str
hub_model_id: str = ""
model_preset: str = ""
model_name: str = ""
num_epochs: int
all_branches: bool
push_to_hub: bool
output_subdir: str
continue_from_latest_artifact: bool = True
continue_model_path: str = ""
@model_validator(mode="after")
def validate_model_target(self) -> HumanTrainingLaunchSpec:
resolved_model_repo = self.hub_model_id or self.model_repo
if not resolved_model_repo:
raise ValueError("Jānorāda hub_model_id vai model_repo.")
self.hub_model_id = resolved_model_repo
self.model_repo = resolved_model_repo
return self
def resolve_human_training_stage_dir(persistent_dir: str, run_id: str) -> Path:
root = Path(persistent_dir).expanduser().resolve()
target = (root / HUMAN_TRAINING_STAGE_DIRNAME / run_id).resolve()
if os.path.commonpath([str(root), str(target)]) != str(root):
raise ValueError(
"Human training staging direktorijai jāatrodas persistent storage ietvaros."
)
return target
def stage_human_training_artifacts(
request: HumanTrainingRequest,
*,
persistent_dir: str,
) -> dict[str, Any]:
run_id = _new_run_id()
stage_dir = resolve_human_training_stage_dir(persistent_dir, run_id)
stage_dir.mkdir(parents=True, exist_ok=True)
train_records = _build_train_records(request)
eval_records = _build_eval_records(request)
preference_dataset = _build_preference_dataset(request)
quality_config = DatasetQualityGateConfig(min_text_chars=HUMAN_TRAINING_MIN_TEXT_CHARS)
filtered_train, train_report = apply_quality_gate_to_records(
train_records,
split_name="train",
config=quality_config,
)
filtered_eval, eval_report = (
apply_quality_gate_to_records(
eval_records,
split_name="eval",
config=quality_config,
)
if eval_records
else ([], None)
)
quality_report = build_dataset_quality_report(
config=quality_config,
train_report=train_report,
eval_report=eval_report,
).to_dict()
preference_examples = [PreferenceExample(**item) for item in preference_dataset["preferences"]]
preference_summary = (
summarize_preference_dataset(preference_examples) if preference_examples else None
)
blind_review = (
build_blind_side_by_side_artifact(
preference_examples,
seed=HUMAN_TRAINING_BLIND_REVIEW_SEED,
)
if preference_examples
else None
)
human_eval_summary = (
build_human_eval_summary(preference_examples) if preference_examples else None
)
artifact_specs: dict[str, dict[str, Any]] = {
"train_dataset": {
"repo_path": f"data/{HUMAN_TRAINING_REPO_PREFIX}/{run_id}/train.json",
"payload": filtered_train,
"record_count": len(filtered_train),
},
"dataset_quality_report": {
"repo_path": f"artifacts/{HUMAN_TRAINING_REPO_PREFIX}/{run_id}/dataset-quality-report.json",
"payload": quality_report,
"record_count": quality_report["kept_records"],
},
}
if filtered_eval:
artifact_specs["eval_dataset"] = {
"repo_path": f"artifacts/{HUMAN_TRAINING_REPO_PREFIX}/{run_id}/eval.json",
"payload": filtered_eval,
"record_count": len(filtered_eval),
}
if preference_dataset["preferences"]:
artifact_specs["preference_dataset"] = {
"repo_path": f"artifacts/{HUMAN_TRAINING_REPO_PREFIX}/{run_id}/preferences.json",
"payload": preference_dataset,
"record_count": len(preference_dataset["preferences"]),
}
if preference_summary is not None:
artifact_specs["preference_summary"] = {
"repo_path": f"artifacts/{HUMAN_TRAINING_REPO_PREFIX}/{run_id}/preference-summary.json",
"payload": preference_summary,
"record_count": preference_summary["total_examples"],
}
if blind_review is not None:
artifact_specs["blind_review"] = {
"repo_path": f"artifacts/{HUMAN_TRAINING_REPO_PREFIX}/{run_id}/blind-review.json",
"payload": blind_review,
"record_count": blind_review["total_pairs"],
}
if human_eval_summary is not None:
artifact_specs["human_eval_summary"] = {
"repo_path": f"artifacts/{HUMAN_TRAINING_REPO_PREFIX}/{run_id}/human-eval-summary.json",
"payload": human_eval_summary,
"record_count": human_eval_summary["total_examples"],
}
artifacts: dict[str, dict[str, Any]] = {}
for name, spec in artifact_specs.items():
local_path = stage_dir / Path(spec["repo_path"]).name
_write_json(local_path, spec["payload"])
artifacts[name] = {
"local_path": str(local_path),
"repo_path": spec["repo_path"],
"record_count": spec["record_count"],
}
launch_spec = HumanTrainingLaunchSpec(
dataset_repo=request.dataset_repo,
model_repo=request.model_repo,
hub_model_id=request.hub_model_id,
model_preset="" if request.model_name else request.model_preset,
model_name=request.model_name,
num_epochs=request.num_epochs,
all_branches=request.all_branches,
push_to_hub=request.push_to_hub,
output_subdir=request.output_subdir,
continue_from_latest_artifact=request.continue_from_latest_artifact,
continue_model_path=request.continue_model_path,
)
manifest = {
"artifact_type": "human-training-manifest",
"run_id": run_id,
"staged_at": _timestamp(),
"stage_dir": str(stage_dir),
"dataset_repo": request.dataset_repo,
"model_repo": request.model_repo,
"input_summary": {
"profile_facts": len(request.profile_facts),
"profile_preferences": len(request.profile_preferences),
"response_instructions": len(request.response_instructions),
"conversation_examples": len(request.conversation_examples),
"preference_pairs": len(request.preference_pairs),
"eval_examples": len(request.eval_examples),
},
"quality_report": quality_report,
"artifacts": artifacts,
"training_request": launch_spec.model_dump(),
"ready_for_review": True,
"ready_for_training": bool(filtered_train),
}
_write_json(stage_dir / "manifest.json", manifest)
return manifest
def load_human_training_manifest(persistent_dir: str, run_id: str) -> dict[str, Any]:
manifest_path = resolve_human_training_stage_dir(persistent_dir, run_id) / "manifest.json"
if not manifest_path.is_file():
raise FileNotFoundError(f"Human training staging ieraksts {run_id} nav atrasts.")
return json.loads(manifest_path.read_text(encoding="utf-8"))
def publish_human_training_artifacts(
manifest: dict[str, Any],
*,
save_file: Any,
) -> list[dict[str, Any]]:
published: list[dict[str, Any]] = []
dataset_repo = str(manifest["dataset_repo"])
run_id = str(manifest["run_id"])
for artifact_name, artifact in manifest.get("artifacts", {}).items():
repo_path = artifact.get("repo_path")
local_path = artifact.get("local_path")
if not isinstance(repo_path, str) or not isinstance(local_path, str):
continue
content = Path(local_path).read_text(encoding="utf-8")
result = save_file(
repo_id=dataset_repo,
repo_type="dataset",
path_in_repo=repo_path,
content=content,
commit_message=f"Add human training artifacts for {run_id}",
)
published.append({"artifact": artifact_name, **result})
return published
def build_human_training_launch_spec(manifest: dict[str, Any]) -> HumanTrainingLaunchSpec:
return HumanTrainingLaunchSpec.model_validate(manifest["training_request"])
def _build_train_records(request: HumanTrainingRequest) -> list[dict[str, Any]]:
records: list[dict[str, Any]] = []
profile_record = _build_profile_record(request)
if profile_record is not None:
records.append(profile_record)
for example in request.conversation_examples:
record = {
"user": example.user,
"assistant": example.assistant,
"metadata": {"source": "human_training_space"},
}
if example.context:
record["context"] = example.context
records.append(record)
return records
def _build_profile_record(request: HumanTrainingRequest) -> dict[str, Any] | None:
sections: list[str] = []
if request.profile_facts:
sections.append(
"Fakti par lietotāju:\n" + "\n".join(f"- {item}" for item in request.profile_facts)
)
if request.profile_preferences:
sections.append(
"Lietotāja preferences:\n"
+ "\n".join(f"- {item}" for item in request.profile_preferences)
)
if request.response_instructions:
sections.append(
"Atbildēšanas instrukcijas:\n"
+ "\n".join(f"- {item}" for item in request.response_instructions)
)
if not sections:
return None
return {
"prompt": "Iegaumē šo lietotāja profilu un atbildē saskaņā ar to turpmākajās sarunās.",
"completion": "\n\n".join(sections),
"metadata": {
"source": "human_training_space",
"artifact": "profile_memory",
},
}
def _build_eval_records(request: HumanTrainingRequest) -> list[dict[str, Any]]:
records: list[dict[str, Any]] = []
for example in request.eval_examples:
record = {
"prompt": example.prompt,
"completion": example.completion,
"metadata": {"source": "human_training_space", "artifact": "eval"},
}
if example.context:
record["context"] = example.context
records.append(record)
return records
def _build_preference_dataset(request: HumanTrainingRequest) -> dict[str, Any]:
preferences: list[dict[str, Any]] = []
for index, pair in enumerate(request.preference_pairs, start=1):
item = {
"prompt": pair.prompt,
"chosen": pair.chosen,
"rejected": pair.rejected,
"context": pair.context,
"source": "human_training_space",
"source_type": "real_reviewer",
"reviewer_segment": "self-training",
"preference_outcome": "chosen",
"confidence": pair.confidence,
"pair_id": f"human-training-{index:04d}",
"blind": True,
"production_like": True,
"tags": ["human-training"],
}
preferences.append(item)
return {
"artifact_type": "human-training-preferences",
"preferences": preferences,
}
def _write_json(path: Path, payload: Any) -> None:
path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
|