rework dataset function and class structure

This commit is contained in:
2026-04-05 02:21:15 +02:00
parent 066a93f755
commit 332e86e3cc
9 changed files with 318 additions and 248 deletions
+9 -54
View File
@@ -1,6 +1,8 @@
import argparse, hashlib, shutil, glob, json
import argparse, hashlib, shutil, json
from pathlib import Path
from server.dataset.DatasetIO import DatasetIO
class DatasetCurator:
def __init__(self, input_files:list[str], output_file:str, min_turn:int=6, late_turn:int=20, max_safe_options:int=2, min_score:int=3, append:bool=False, archive_input:bool=False, archive_dir:str|None=None):
self.input_files = input_files
@@ -11,45 +13,7 @@ class DatasetCurator:
self.min_score = min_score
self.append = append
self.archive_input = archive_input
self.archive_dir = (
Path(archive_dir) if archive_dir else self.output_file.parent / "archive"
)
def _resolve_input_files(self):
resolved = []
seen = set()
for item in self.input_files:
path = Path(item)
if path.is_dir():
for file_path in sorted(path.rglob("*.jsonl")):
key = str(file_path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(file_path)
continue
if any(ch in item for ch in "*?[]"):
for match in sorted(glob.glob(item)):
file_path = Path(match)
if not file_path.is_file():
continue
key = str(file_path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(file_path)
continue
if path.is_file():
key = str(path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(path)
return resolved
self.archive_dir = Path(archive_dir) if archive_dir else self.output_file.parent / "archive"
def _safe_options_count(self, row:dict):
history = row.get("history", {})
@@ -98,7 +62,7 @@ class DatasetCurator:
def curate(self):
self.output_file.parent.mkdir(parents=True, exist_ok=True)
input_paths = self._resolve_input_files()
input_paths = DatasetIO.resolve_input_files(self.input_files)
total = 0
kept = 0
@@ -108,24 +72,15 @@ class DatasetCurator:
seen_states = set()
if self.append and self.output_file.exists():
with self.output_file.open("r", encoding="utf-8") as existing:
for line in existing:
if not line.strip():
continue
row = json.loads(line)
for row in DatasetIO.iter_jsonl_rows(self.output_file):
state_key = self._state_hash(row)
seen_states.add((state_key, row.get("move")))
mode = "a" if self.append else "w"
with self.output_file.open(mode, encoding="utf-8") as dst:
with DatasetIO.open_text(self.output_file, mode) as dst:
for input_path in input_paths:
with input_path.open("r", encoding="utf-8") as src:
for line in src:
if not line.strip():
continue
for row in DatasetIO.iter_jsonl_rows(input_path):
total += 1
row = json.loads(line)
if not row.get("is_good_move", False):
skipped_quality += 1
@@ -220,7 +175,7 @@ if __name__ == "__main__":
"--input",
action="append",
required=True,
help="Input JSONL file, directory, or glob pattern. Repeat for multiple inputs.",
help="Input JSONL/JSONL.GZ file, directory, or glob pattern. Repeat for multiple inputs.",
)
parser.add_argument("--output", required=True, help="Output JSONL file")
parser.add_argument("--min-turn", type=int, default=6)
+4 -4
View File
@@ -1,15 +1,15 @@
from pathlib import Path
import argparse, json
from server.dataset.DatasetIO import DatasetIO
class DatasetExporter:
def __init__(self, input_dir:str, output_file:str):
self.input_dir = Path(input_dir)
self.output_file = Path(output_file)
def _iter_game_files(self):
if not self.input_dir.exists():
return []
return sorted(self.input_dir.rglob("*.json"))
return DatasetIO.list_directory_files(self.input_dir, directory_pattern="*.json")
def _extract_samples(self, payload:dict, source_file:Path):
dataset = payload.get("dataset", {})
@@ -37,7 +37,7 @@ class DatasetExporter:
self.output_file.parent.mkdir(parents=True, exist_ok=True)
sample_count = 0
with self.output_file.open("w", encoding="utf-8") as output:
with DatasetIO.open_text(self.output_file, "w") as output:
for game_file in game_files:
with game_file.open("r", encoding="utf-8") as source:
payload = json.load(source)
+134
View File
@@ -0,0 +1,134 @@
from typing import Any, TextIO, cast
from datetime import UTC, datetime
from pathlib import Path
import gzip, glob, json
class DatasetIO:
@staticmethod
def resolve_input_files(input_files: list[str], directory_pattern: str | tuple[str, ...] = ("*.jsonl", "*.jsonl.gz")) -> list[Path]:
patterns = (
(directory_pattern,)
if isinstance(directory_pattern, str)
else directory_pattern
)
resolved = []
seen = set()
for item in input_files:
path = Path(item)
if path.is_dir():
for pattern in patterns:
for file_path in sorted(path.rglob(pattern)):
key = str(file_path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(file_path)
continue
if any(ch in item for ch in "*?[]"):
for match in sorted(glob.glob(item)):
file_path = Path(match)
if not file_path.is_file():
continue
key = str(file_path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(file_path)
continue
if path.is_file():
key = str(path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(path)
return resolved
@staticmethod
def list_directory_files(input_dir:Path, directory_pattern:str) -> list[Path]:
if not input_dir.exists():
return []
return sorted(input_dir.rglob(directory_pattern))
@staticmethod
def open_text(path:Path, mode:str="r"):
text_mode = mode if "t" in mode else f"{mode}t"
open_fn = gzip.open if path.suffix == ".gz" else open
return cast(TextIO, open_fn(path, text_mode, encoding="utf-8"))
@staticmethod
def iter_jsonl_rows(path:Path):
with DatasetIO.open_text(path, "r") as handle:
for line in handle:
if line.strip():
yield json.loads(line)
@staticmethod
def append_jsonl_row(path:Path, row:dict[str, Any]):
with DatasetIO.open_text(path, "a") as raw_handle:
handle = cast(TextIO, raw_handle)
handle.write(json.dumps(row, ensure_ascii=False) + "\n")
@staticmethod
def count_jsonl_rows(path:Path) -> int:
count = 0
candidates = [path]
if path.suffix == ".gz":
candidates.append(path.with_suffix(""))
else:
candidates.append(Path(f"{path}.gz"))
seen = set()
for candidate in candidates:
if candidate in seen:
continue
seen.add(candidate)
if not candidate.exists() or not candidate.is_file():
continue
try:
with DatasetIO.open_text(candidate, "r") as handle:
for line in handle:
if line.strip():
count += 1
except (OSError, UnicodeError):
continue
return count
@staticmethod
def rotate_and_gzip_if_size_reached(path:Path, max_bytes:int) -> bool:
if max_bytes <= 0:
return False
if not path.exists() or not path.is_file():
return False
if path.stat().st_size < max_bytes:
return False
timestamp = datetime.now(UTC).strftime("%Y%m%d-%H%M%S")
if path.suffix == ".jsonl":
rotated_name = f"{path.stem}.{timestamp}.jsonl"
else:
rotated_name = f"{path.name}.{timestamp}"
rotated_path = path.with_name(rotated_name)
suffix = 1
while rotated_path.exists():
suffix += 1
if path.suffix == ".jsonl":
rotated_name = f"{path.stem}.{timestamp}.{suffix}.jsonl"
else:
rotated_name = f"{path.name}.{timestamp}.{suffix}"
rotated_path = path.with_name(rotated_name)
path.rename(rotated_path)
with rotated_path.open("rb") as src:
with gzip.open(f"{rotated_path}.gz", "wb") as dst:
dst.writelines(src)
rotated_path.unlink()
return True
+6 -45
View File
@@ -1,50 +1,16 @@
from collections import Counter, defaultdict
import argparse, glob, json, re
import argparse, json, re
from datetime import datetime
from pathlib import Path
from server.dataset.DatasetIO import DatasetIO
class DatasetStats:
DAY_PATTERN = re.compile(r"(\d{4}-\d{2}-\d{2})")
def __init__(self, input_files:list[str]):
self.input_files = input_files
def _resolve_input_files(self):
resolved = []
seen = set()
for item in self.input_files:
path = Path(item)
if path.is_dir():
for file_path in sorted(path.rglob("*.jsonl")):
key = str(file_path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(file_path)
continue
if any(ch in item for ch in "*?[]"):
for match in sorted(glob.glob(item)):
file_path = Path(match)
if not file_path.is_file():
continue
key = str(file_path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(file_path)
continue
if path.is_file():
key = str(path.resolve())
if key in seen:
continue
seen.add(key)
resolved.append(path)
return resolved
def _infer_day(self, file_path:Path):
match = self.DAY_PATTERN.search(file_path.name)
if match:
@@ -83,7 +49,7 @@ class DatasetStats:
return None
def analyze(self):
files = self._resolve_input_files()
files = DatasetIO.resolve_input_files(self.input_files)
totals = {
"rows": 0,
@@ -99,12 +65,7 @@ class DatasetStats:
for file_path in files:
day = self._infer_day(file_path)
with file_path.open("r", encoding="utf-8") as source:
for line in source:
if not line.strip():
continue
row = json.loads(line)
for row in DatasetIO.iter_jsonl_rows(file_path):
game_id = row.get("game_id")
if not game_id:
continue
@@ -225,7 +186,7 @@ if __name__ == "__main__":
"--input",
action="append",
required=True,
help="Input JSONL file, directory, or glob pattern. Repeat for multiple inputs.",
help="Input JSONL/JSONL.GZ file, directory, or glob pattern. Repeat for multiple inputs.",
)
parser.add_argument(
"--output",
+7 -69
View File
@@ -1,8 +1,8 @@
from datetime import UTC, datetime
from pathlib import Path
from typing import Any
import os
import gzip, json, os
from server.dataset.DatasetIO import DatasetIO
class RLBootstrapDataset:
def __init__(self):
@@ -30,76 +30,15 @@ class RLBootstrapDataset:
except ValueError:
return default
@staticmethod
def count_jsonl_rows(path:Path) -> int:
count = 0
candidates = [path]
if path.suffix == ".gz":
candidates.append(path.with_suffix(""))
else:
candidates.append(Path(f"{path}.gz"))
seen = set()
for candidate in candidates:
if candidate in seen:
continue
seen.add(candidate)
if not candidate.exists() or not candidate.is_file():
continue
open_fn = gzip.open if candidate.suffix == ".gz" else open
try:
with open_fn(candidate, "rt", encoding="utf-8") as handle:
for line in handle:
if line.strip():
count += 1
except (OSError, UnicodeError):
continue
return count
@staticmethod
def rotate_and_gzip_if_size_reached(path:Path, max_bytes:int) -> bool:
if max_bytes <= 0:
return False
if not path.exists() or not path.is_file():
return False
if path.stat().st_size < max_bytes:
return False
timestamp = datetime.now(UTC).strftime("%Y%m%d-%H%M%S")
if path.suffix == ".jsonl":
rotated_name = f"{path.stem}.{timestamp}.jsonl"
else:
rotated_name = f"{path.name}.{timestamp}"
rotated_path = path.with_name(rotated_name)
suffix = 1
while rotated_path.exists():
suffix += 1
if path.suffix == ".jsonl":
rotated_name = f"{path.stem}.{timestamp}.{suffix}.jsonl"
else:
rotated_name = f"{path.name}.{timestamp}.{suffix}"
rotated_path = path.with_name(rotated_name)
path.rename(rotated_path)
with rotated_path.open("rb") as src:
with gzip.open(f"{rotated_path}.gz", "wb") as dst:
dst.writelines(src)
rotated_path.unlink()
return True
def refresh_state(self):
if not self.enabled:
self.needs_more_data = False
return
base_rows = self.count_jsonl_rows(self.base_dataset_path)
base_rows = DatasetIO.count_jsonl_rows(self.base_dataset_path)
self.needs_more_data = base_rows < self.min_base_rows
def record_sample(self, game_data:Any, move:str, safe_moves:dict[str, dict[str, int]], reason:str, scores:dict[str, float]|None=None,):
def record_sample(self, game_data:Any, move:str, safe_moves:dict[str, dict[str, int]], reason:str, scores:dict[str, float]|None=None):
if not self.enabled or not self.needs_more_data:
return
@@ -117,8 +56,7 @@ class RLBootstrapDataset:
if scores:
row["scores"] = {k: round(v, 5) for k, v in scores.items()}
with self.output_path.open("a", encoding="utf-8") as handle:
handle.write(json.dumps(row, ensure_ascii=False) + "\n")
self.rotate_and_gzip_if_size_reached(self.output_path, self.max_bytes)
DatasetIO.append_jsonl_row(self.output_path, row)
DatasetIO.rotate_and_gzip_if_size_reached(self.output_path, self.max_bytes)
except Exception:
return
+1
View File
@@ -1,4 +1,5 @@
from .Dataset import Dataset
from .DatasetIO import DatasetIO
from .DatasetExporter import DatasetExporter
from .DatasetCurator import DatasetCurator
from .DatasetStats import DatasetStats
+38 -2
View File
@@ -1,6 +1,6 @@
import json
import tempfile
import unittest
import tempfile, json, gzip
from pathlib import Path
from server.dataset.DatasetExporter import DatasetExporter
@@ -43,5 +43,41 @@ class TestDatasetExporter(unittest.TestCase):
self.assertEqual(first["move"], "up")
self.assertTrue(first["is_good_move"])
def test_export_jsonl_gz(self):
with tempfile.TemporaryDirectory() as tmp:
input_dir = Path(tmp) / "data"
output_file = Path(tmp) / "out" / "dataset.jsonl.gz"
game_file = input_dir / "game-1.json"
game_file.parent.mkdir(parents=True, exist_ok=True)
game_payload = {
"dataset": {
"game": {"id": "g-1", "map": "standard", "type": {"name": "duel"}},
"snake": {"type": "BestBattleSnake"},
"samples": [
{
"turn": 1,
"move": "up",
"is_good_move": True,
"game_board": {"width": 11, "height": 11},
"history": {"data": []},
}
],
}
}
game_file.write_text(json.dumps(game_payload), encoding="utf-8")
report = DatasetExporter(str(input_dir), str(output_file)).export_jsonl()
self.assertEqual(report["games_scanned"], 1)
self.assertEqual(report["samples_exported"], 1)
self.assertTrue(output_file.exists())
with gzip.open(output_file, "rt", encoding="utf-8") as handle:
lines = handle.read().strip().splitlines()
self.assertEqual(len(lines), 1)
first = json.loads(lines[0])
self.assertEqual(first["game_id"], "g-1")
if __name__ == "__main__":
unittest.main()
+46
View File
@@ -0,0 +1,46 @@
import unittest
import tempfile, gzip, json
from pathlib import Path
from server.dataset.DatasetIO import DatasetIO
class TestDatasetIO(unittest.TestCase):
def test_resolve_input_files_includes_jsonl_and_gz(self):
with tempfile.TemporaryDirectory() as tmp:
base = Path(tmp)
(base / "a.jsonl").write_text('{"x":1}\n', encoding="utf-8")
with gzip.open(base / "b.jsonl.gz", "wt", encoding="utf-8") as handle:
handle.write('{"x":2}\n')
paths = DatasetIO.resolve_input_files([str(base)])
names = {path.name for path in paths}
self.assertIn("a.jsonl", names)
self.assertIn("b.jsonl.gz", names)
def test_iter_jsonl_rows_reads_gz_file(self):
with tempfile.TemporaryDirectory() as tmp:
path = Path(tmp) / "rows.jsonl.gz"
with gzip.open(path, "wt", encoding="utf-8") as handle:
handle.write('{"turn":1}\n')
handle.write("\n")
handle.write('{"turn":2}\n')
rows = list(DatasetIO.iter_jsonl_rows(path))
self.assertEqual(len(rows), 2)
self.assertEqual(rows[0]["turn"], 1)
self.assertEqual(rows[1]["turn"], 2)
def test_append_jsonl_row_supports_gz_output(self):
with tempfile.TemporaryDirectory() as tmp:
path = Path(tmp) / "append.jsonl.gz"
DatasetIO.append_jsonl_row(path, {"move": "up"})
DatasetIO.append_jsonl_row(path, {"move": "left"})
with gzip.open(path, "rt", encoding="utf-8") as handle:
lines = handle.read().strip().splitlines()
self.assertEqual(len(lines), 2)
self.assertEqual(json.loads(lines[0])["move"], "up")
if __name__ == "__main__":
unittest.main()
+4 -5
View File
@@ -5,6 +5,7 @@ from pathlib import Path
import tempfile, gzip
from server.dataset.RLBootstrapDataset import RLBootstrapDataset
from server.dataset.DatasetIO import DatasetIO
class TestRLBootstrapDataset(unittest.TestCase):
def test_count_jsonl_rows_reads_gzip_dataset(self):
@@ -15,17 +16,15 @@ class TestRLBootstrapDataset(unittest.TestCase):
handle.write("\n")
handle.write('{"turn":2}\n')
self.assertEqual(RLBootstrapDataset.count_jsonl_rows(dataset_path), 2)
self.assertEqual(RLBootstrapDataset.count_jsonl_rows(Path(tmp) / "base.jsonl"), 2)
self.assertEqual(DatasetIO.count_jsonl_rows(dataset_path), 2)
self.assertEqual(DatasetIO.count_jsonl_rows(Path(tmp) / "base.jsonl"), 2)
def test_rotate_and_gzip_if_size_reached_rotates_jsonl(self):
with tempfile.TemporaryDirectory() as tmp:
path = Path(tmp) / "rl_bootstrap.jsonl"
path.write_text("x" * 200, encoding="utf-8")
rotated = RLBootstrapDataset.rotate_and_gzip_if_size_reached(
path, max_bytes=50
)
rotated = DatasetIO.rotate_and_gzip_if_size_reached(path, max_bytes=50)
self.assertTrue(rotated)
self.assertFalse(path.exists())