File size: 11,345 Bytes
4949db9 | 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 | """ms-swift JSONL conversion pipeline: write, split, validate, register."""
from __future__ import annotations
import argparse
import datetime
import json
import logging
import os
import random
from pathlib import Path
from typing import Any
from judge_training.data.build_records_from_db import extract_model_and_source
from judge_training.data.naming import REGISTRY_PATH, swift_test_path
from judge_training.data.sample import ALLOWED_KEYS, TrainingSample
from judge_training.data.schema import resolve_schema
logger = logging.getLogger(__name__)
def write_swift_jsonl(samples: list[TrainingSample], output_path: str) -> int:
"""Write validated samples to a Swift JSONL file."""
output = Path(output_path)
output.parent.mkdir(parents=True, exist_ok=True)
with open(output, "w", encoding="utf-8") as f:
for sample in samples:
f.write(sample.to_jsonl() + "\n")
logger.info("Wrote %d lines to %s", len(samples), output_path)
return len(samples)
def split_samples_by_video(
samples: list[TrainingSample],
*,
val_ratio: float,
seed: int,
) -> tuple[list[TrainingSample], list[TrainingSample]]:
"""Split samples by video path so all prompts for one video stay together."""
if not 0 < val_ratio < 1:
raise ValueError("--val-ratio must be between 0 and 1")
video_to_samples: dict[str, list[TrainingSample]] = {}
for sample in samples:
video_to_samples.setdefault(sample.video_path, []).append(sample)
video_paths = sorted(video_to_samples)
if not video_paths:
return [], []
rng = random.Random(seed)
rng.shuffle(video_paths)
n_val = max(1, int(len(video_paths) * val_ratio))
if len(video_paths) > 1:
n_val = min(n_val, len(video_paths) - 1)
val_videos = set(video_paths[:n_val])
train_samples: list[TrainingSample] = []
val_samples: list[TrainingSample] = []
for video_path in video_paths:
if video_path in val_videos:
val_samples.extend(video_to_samples[video_path])
else:
train_samples.extend(video_to_samples[video_path])
logger.info(
"Train/val split: %d videos train, %d videos val (%d/%d samples)",
len(video_paths) - n_val,
n_val,
len(train_samples),
len(val_samples),
)
return train_samples, val_samples
def validate_jsonl(jsonl_path: str) -> dict[str, Any]:
"""Validate a Swift JSONL file by parsing each line into TrainingSample."""
errors: list[str] = []
total = 0
valid = 0
score_dist: dict[int, int] = {}
key_dist: dict[str, int] = {}
video_model_dist: dict[str, int] = {}
with open(jsonl_path) as f:
for line_num, line in enumerate(f, 1):
line = line.strip()
if not line:
continue
total += 1
try:
row = json.loads(line)
except json.JSONDecodeError as e:
errors.append(f"Line {line_num}: invalid JSON: {e}")
continue
try:
sample = TrainingSample.from_jsonl_row(row)
except ValueError as e:
errors.append(f"Line {line_num}: {e}")
continue
valid += 1
parsed = json.loads(sample.assistant)
for k, v in parsed.items():
if k in ALLOWED_KEYS:
key_dist[k] = key_dist.get(k, 0) + 1
score_dist[v] = score_dist.get(v, 0) + 1
parts = Path(sample.video_path).parts
for i, part in enumerate(parts):
if part == "videos" and i + 1 < len(parts):
model, _ = extract_model_and_source(parts[i + 1])
video_model_dist[model] = video_model_dist.get(model, 0) + 1
break
summary: dict[str, Any] = {
"file": jsonl_path,
"total_lines": total,
"valid_lines": valid,
"errors": errors[:20],
}
print(f"\n=== Validation: {jsonl_path} ===")
print(f" Total lines: {total}")
print(f" Valid: {valid}")
if valid > 0:
print("\n Score distribution:")
for score in sorted(score_dist):
print(f" {score}: {score_dist[score]}")
print("\n Key distribution:")
for key in sorted(key_dist):
print(f" {key}: {key_dist[key]}")
if video_model_dist:
print("\n Video model distribution:")
for model in sorted(video_model_dist):
print(f" {model}: {video_model_dist[model]}")
if errors:
print(f" Errors ({len(errors)} total, showing first 10):")
for error in errors[:10]:
print(f" {error}")
else:
print(" No errors found.")
return summary
def _is_ephemeral_output(output: str) -> bool:
"""Detect debug/temp output paths that shouldn't be registered."""
resolved = str(Path(output).resolve())
if resolved.startswith("/tmp") or resolved.startswith("/var/tmp"):
return True
return False
def write_test_splits(
splits: dict[str, list[TrainingSample]],
stem: str,
) -> None:
"""Write and validate test-split JSONL files."""
for split_name in ("test_prompt", "test_model", "test_both"):
samples = splits.get(split_name, [])
if not samples:
logger.info(" %s: empty, skipping", split_name)
continue
path = swift_test_path(stem, split_name)
write_swift_jsonl(samples, path)
validate_jsonl(path)
def append_training_registry(
*,
module: str,
output: str,
holdout_model: str | None,
val_output: str | None,
val_ratio: float | None,
seed: int | None,
n_train: int,
n_val: int | None,
metadata: dict[str, Any],
command_args: list[str],
) -> None:
"""Append one run entry to training_registry.json.
Auto-skips /tmp outputs. If output_train already exists in the
registry, the old entry is replaced instead of duplicated.
"""
if _is_ephemeral_output(output):
logger.info("Output is in /tmp — skipping training registry")
return
if REGISTRY_PATH.is_file():
with open(REGISTRY_PATH) as f:
registry = json.load(f)
else:
REGISTRY_PATH.parent.mkdir(parents=True, exist_ok=True)
registry = []
prompt_config = metadata.get("prompt_config") or "default.yaml"
schema_info = resolve_schema(prompt_config)
entry = {
"scheme": schema_info.scheme,
"schema": schema_info.version,
"schema_fingerprint": schema_info.fingerprint,
"prompt_config": prompt_config,
"prompt_config_source": metadata.get("prompt_config_source", "unknown"),
"label_source": metadata.get("label_source", "unknown"),
"target_format": metadata.get("target_format", "unknown"),
"dims": metadata.get("dims", []),
"score_scale": metadata.get("score_scale", "unknown"),
"holdout_model": holdout_model,
"val_ratio": val_ratio if val_output else None,
"seed": seed if val_output else None,
"n_train": n_train,
"n_val": n_val,
"output_train": output,
"output_val": val_output,
"datetime": datetime.datetime.now().strftime("%m-%d %H:%M:%S"),
"gen_command": " ".join(
["python", "-m", module, "convert", *command_args]
),
}
replaced = False
for i, existing in enumerate(registry):
if existing.get("output_train") == output:
registry[i] = entry
replaced = True
break
if not replaced:
registry.append(entry)
with open(REGISTRY_PATH, "w") as f:
json.dump(registry, f, indent=2, ensure_ascii=False)
f.write("\n")
action = "Replaced" if replaced else "Appended"
logger.info("%s training registry: scheme=%s, schema=%s, output=%s", action, schema_info.scheme, schema_info.version, output)
def validate_cli(jsonl_path: str) -> int:
"""Shared implementation for validate subcommands."""
if not os.path.isfile(jsonl_path):
logger.error("File not found: %s", jsonl_path)
return 1
summary = validate_jsonl(jsonl_path)
if summary["total_lines"] == 0:
logger.error("Empty file: %s", jsonl_path)
return 1
return 0 if not summary["errors"] else 1
def add_common_convert_args(parser: argparse.ArgumentParser) -> None:
"""Add shared convert subparser arguments (base_dir, holdout, val split)."""
parser.add_argument(
"--base_dir",
default=".",
help="Base directory for resolving video paths",
)
parser.add_argument(
"--holdout_model",
default="veo-3.1",
help="Video model to hold out for testing",
)
parser.add_argument(
"--val-output",
default=None,
help="Output path for validation split JSONL",
)
parser.add_argument(
"--val-ratio",
type=float,
default=0.05,
help="Fraction of videos to hold out for validation (default: 0.05)",
)
parser.add_argument(
"--seed",
type=int,
default=42,
help="Random seed for train/val split (default: 42)",
)
parser.add_argument(
"--skip-registry",
action="store_true",
help="Write/validate JSONL but do not append docs/exp-results/training/training_registry.json",
)
def append_val_command_args(parts: list[str], args: argparse.Namespace) -> None:
"""Append shared val-split CLI args to a command_args list."""
if args.base_dir != ".":
parts.extend(["--base_dir", args.base_dir])
if args.holdout_model:
parts.extend(["--holdout_model", args.holdout_model])
if args.val_output:
parts.extend([
"--val-output", args.val_output,
"--val-ratio", str(args.val_ratio),
"--seed", str(args.seed),
])
def run_convert(
*,
samples: list[TrainingSample],
args: argparse.Namespace,
metadata: dict[str, Any],
module: str,
command_args: list[str],
) -> int:
"""Shared convert flow: split, write, validate, register."""
if not samples:
raise ValueError("No valid samples; check input paths and --base_dir")
n_val_written = None
if args.val_output:
train_samples, val_samples = split_samples_by_video(
samples,
val_ratio=args.val_ratio,
seed=args.seed,
)
n_val_written = write_swift_jsonl(val_samples, args.val_output)
samples = train_samples
n_train_written = write_swift_jsonl(samples, args.output)
summary = validate_jsonl(args.output)
if n_val_written is not None:
validate_jsonl(args.val_output)
if args.skip_registry:
logger.info("Skipping training registry append")
else:
append_training_registry(
module=module,
output=args.output,
holdout_model=args.holdout_model,
val_output=args.val_output,
val_ratio=args.val_ratio,
seed=args.seed,
n_train=n_train_written,
n_val=n_val_written,
metadata=metadata,
command_args=command_args,
)
return 0 if not summary["errors"] else 1
|