Pipeline code snapshot (rubric_gen + gke) matching commit bee1bd2

f45837e verified 2 months ago

6.34 kB

#!/usr/bin/env python3
"""A temporary split over only what is already on disk, weighted toward EVAL.
    python3 make_fast_split.py --out split_fast.json
    RUBRIC_GEN_SPLIT=$PWD/split_fast.json python3 run.py
WHY
The real manifest is 15 train / 60 eval, which is the right design. What
actually finished downloading is the opposite shape -- 10 train against 6 eval
-- because the train competitions happen to be the enormous ones (seti 245 GB,
herbarium 149 GB) while the eval ones are small. Measuring against 6 held-out
competitions when 10 are available for mining is backwards: rubric quality is
bounded by how much held-out signal we can measure, not by how much we mined.
So this builds a 4 train / 12 eval split over exactly the competitions on disk,
to get a real held-out number today instead of after another ~500 GB of
downloads.
THE CORRECTNESS CATCH
Competitions already mined under the real split CANNOT simply be relabelled as
eval. Rubrics exist that were derived from them, and scoring those rubrics on
their own source data reports memorization as transfer. The only sound way to
shrink the train side is to shrink it and REGENERATE -- so `--train` picks the
new source set, and everything mined previously must be discarded and rebuilt
from just those. The script prints that instruction rather than assuming it.
TRAIN SELECTION
Four competitions spanning four input modalities -- tabular, text, natural
image, medical image. Diversity matters more than count here: rubrics mined
from four different kinds of task generalize better than from eight variants of
one, and the earlier run showed a single competition's idioms dominating the
output when the corpus was narrow.
Its experiment_id is `rubric-gen-fast-signal`, stamped into every artifact,
audit report and result file produced under it -- so a number from this split
can never be quietly presented as a number from the real one.
"""

from __future__ import annotations

import argparse
import json
import sys
from pathlib import Path

HERE = Path(__file__).resolve().parent
sys.path.insert(0, str(HERE))

import split as split_mod

DATA_DIR = Path("/workspace/edwardosunny/mle-bench/data")

# One per modality, all small enough to re-mine quickly.
DEFAULT_TRAIN = [
    "tabular-playground-series-dec-2021",   # tabular
    "spooky-author-identification",         # text classification
    "dog-breed-identification",             # natural image, fine-grained
    "osic-pulmonary-fibrosis-progression",  # medical image + tabular regression
]


def prepared(data_dir: Path, comp: str) -> bool:
    pub = data_dir / comp / "prepared" / "public"
    return pub.is_dir() and any(pub.iterdir())


def main() -> int:
    ap = argparse.ArgumentParser(description=__doc__,
                                 formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--data-dir", default=str(DATA_DIR))
    ap.add_argument("--train", default=",".join(DEFAULT_TRAIN))
    ap.add_argument("--out", default=str(HERE / "split_fast.json"))
    args = ap.parse_args()

    data_dir = Path(args.data_dir)
    real = split_mod.Split(split_mod.load_manifest(HERE / "split_manifest.json"))
    universe = real.source + real.evaluation

    on_disk = [c for c in universe if prepared(data_dir, c)]
    train = [c.strip() for c in args.train.split(",") if c.strip()]

    missing = [c for c in train if c not in on_disk]
    if missing:
        raise SystemExit(f"requested train competitions are not prepared: {missing}")

    evaluation = [c for c in on_disk if c not in set(train)]
    if not evaluation:
        raise SystemExit("no evaluation competitions left on disk")

    manifest = {
        "experiment_id": "rubric-gen-fast-signal",
        # Frozen because it IS fixed once written, and because
        # split.get(strict=True) -- which every mining entry point uses --
        # refuses an unfrozen manifest. Marking it False would have blocked the
        # very run it exists for. The guard against mistaking this for the real
        # thing is the experiment_id, stamped into every artifact and audit.
        "frozen": True,
        "purpose": "Early signal over only the competitions already downloaded. "
                   "Weighted toward eval because held-out measurement, not "
                   "mining volume, is what bounds the result right now.",
        "supersedes_nothing": "The real split is split_manifest.json (15/60). "
                              "This one covers only what is on disk.",
        "selection": {
            "train": "four competitions, one per input modality -- tabular, "
                     "text, natural image, medical image. Diversity beats count: "
                     "a narrow corpus lets one competition's idioms dominate.",
            "eval": "every other prepared competition.",
        },
        "warning": "Rubrics previously mined under the real split were derived "
                   "from competitions that are EVAL here. They must be "
                   "regenerated from this train set before any measurement, or "
                   "the result reports memorization as transfer.",
        "counts": {"train": len(train), "eval": len(evaluation),
                   "train_fraction": round(len(train) / len(on_disk), 4)},
        "source_competitions": [{"id": c} for c in sorted(train)],
        "evaluation_competitions": [{"id": c} for c in sorted(evaluation)],
    }
    Path(args.out).write_text(json.dumps(manifest, indent=2) + "\n")

    print(f"{len(on_disk)} competitions on disk -> {len(train)} train / "
          f"{len(evaluation)} eval "
          f"({100*len(train)/len(on_disk):.0f}% / "
          f"{100*len(evaluation)/len(on_disk):.0f}%)")
    print("\nTRAIN (mine rubrics from these):")
    for c in sorted(train):
        print(f"  {c}")
    print("\nEVAL (measure on these):")
    for c in sorted(evaluation):
        was = "was train in the real split" if real.is_source(c) else ""
        print(f"  {c:52s} {was}")
    print(f"\n-> {args.out}")
    print("\nREGENERATE before measuring -- existing rubrics were mined from "
          "competitions that are EVAL here:")
    print(f"  export RUBRIC_GEN_SPLIT={args.out}")
    print("  rm -f data/states.jsonl data/conceptual_*.jsonl out/*/*.md")
    return 0


if __name__ == "__main__":
    sys.exit(main())

	#!/usr/bin/env python3
	"""A temporary split over only what is already on disk, weighted toward EVAL.

	python3 make_fast_split.py --out split_fast.json
	RUBRIC_GEN_SPLIT=$PWD/split_fast.json python3 run.py

	WHY
	The real manifest is 15 train / 60 eval, which is the right design. What
	actually finished downloading is the opposite shape -- 10 train against 6 eval
	-- because the train competitions happen to be the enormous ones (seti 245 GB,
	herbarium 149 GB) while the eval ones are small. Measuring against 6 held-out
	competitions when 10 are available for mining is backwards: rubric quality is
	bounded by how much held-out signal we can measure, not by how much we mined.

	So this builds a 4 train / 12 eval split over exactly the competitions on disk,
	to get a real held-out number today instead of after another ~500 GB of
	downloads.

	THE CORRECTNESS CATCH
	Competitions already mined under the real split CANNOT simply be relabelled as
	eval. Rubrics exist that were derived from them, and scoring those rubrics on
	their own source data reports memorization as transfer. The only sound way to
	shrink the train side is to shrink it and REGENERATE -- so `--train` picks the
	new source set, and everything mined previously must be discarded and rebuilt
	from just those. The script prints that instruction rather than assuming it.

	TRAIN SELECTION
	Four competitions spanning four input modalities -- tabular, text, natural
	image, medical image. Diversity matters more than count here: rubrics mined
	from four different kinds of task generalize better than from eight variants of
	one, and the earlier run showed a single competition's idioms dominating the
	output when the corpus was narrow.

	Its experiment_id is `rubric-gen-fast-signal`, stamped into every artifact,
	audit report and result file produced under it -- so a number from this split
	can never be quietly presented as a number from the real one.
	"""

	from __future__ import annotations

	import argparse
	import json
	import sys
	from pathlib import Path

	HERE = Path(__file__).resolve().parent
	sys.path.insert(0, str(HERE))

	import split as split_mod

	DATA_DIR = Path("/workspace/edwardosunny/mle-bench/data")

	# One per modality, all small enough to re-mine quickly.
	DEFAULT_TRAIN = [
	"tabular-playground-series-dec-2021", # tabular
	"spooky-author-identification", # text classification
	"dog-breed-identification", # natural image, fine-grained
	"osic-pulmonary-fibrosis-progression", # medical image + tabular regression
	]


	def prepared(data_dir: Path, comp: str) -> bool:
	pub = data_dir / comp / "prepared" / "public"
	return pub.is_dir() and any(pub.iterdir())


	def main() -> int:
	ap = argparse.ArgumentParser(description=__doc__,
	formatter_class=argparse.RawDescriptionHelpFormatter)
	ap.add_argument("--data-dir", default=str(DATA_DIR))
	ap.add_argument("--train", default=",".join(DEFAULT_TRAIN))
	ap.add_argument("--out", default=str(HERE / "split_fast.json"))
	args = ap.parse_args()

	data_dir = Path(args.data_dir)
	real = split_mod.Split(split_mod.load_manifest(HERE / "split_manifest.json"))
	universe = real.source + real.evaluation

	on_disk = [c for c in universe if prepared(data_dir, c)]
	train = [c.strip() for c in args.train.split(",") if c.strip()]

	missing = [c for c in train if c not in on_disk]
	if missing:
	raise SystemExit(f"requested train competitions are not prepared: {missing}")

	evaluation = [c for c in on_disk if c not in set(train)]
	if not evaluation:
	raise SystemExit("no evaluation competitions left on disk")

	manifest = {
	"experiment_id": "rubric-gen-fast-signal",
	# Frozen because it IS fixed once written, and because
	# split.get(strict=True) -- which every mining entry point uses --
	# refuses an unfrozen manifest. Marking it False would have blocked the
	# very run it exists for. The guard against mistaking this for the real
	# thing is the experiment_id, stamped into every artifact and audit.
	"frozen": True,
	"purpose": "Early signal over only the competitions already downloaded. "
	"Weighted toward eval because held-out measurement, not "
	"mining volume, is what bounds the result right now.",
	"supersedes_nothing": "The real split is split_manifest.json (15/60). "
	"This one covers only what is on disk.",
	"selection": {
	"train": "four competitions, one per input modality -- tabular, "
	"text, natural image, medical image. Diversity beats count: "
	"a narrow corpus lets one competition's idioms dominate.",
	"eval": "every other prepared competition.",
	},
	"warning": "Rubrics previously mined under the real split were derived "
	"from competitions that are EVAL here. They must be "
	"regenerated from this train set before any measurement, or "
	"the result reports memorization as transfer.",
	"counts": {"train": len(train), "eval": len(evaluation),
	"train_fraction": round(len(train) / len(on_disk), 4)},
	"source_competitions": [{"id": c} for c in sorted(train)],
	"evaluation_competitions": [{"id": c} for c in sorted(evaluation)],
	}
	Path(args.out).write_text(json.dumps(manifest, indent=2) + "\n")

	print(f"{len(on_disk)} competitions on disk -> {len(train)} train / "
	f"{len(evaluation)} eval "
	f"({100*len(train)/len(on_disk):.0f}% / "
	f"{100*len(evaluation)/len(on_disk):.0f}%)")
	print("\nTRAIN (mine rubrics from these):")
	for c in sorted(train):
	print(f" {c}")
	print("\nEVAL (measure on these):")
	for c in sorted(evaluation):
	was = "was train in the real split" if real.is_source(c) else ""
	print(f" {c:52s} {was}")
	print(f"\n-> {args.out}")
	print("\nREGENERATE before measuring -- existing rubrics were mined from "
	"competitions that are EVAL here:")
	print(f" export RUBRIC_GEN_SPLIT={args.out}")
	print(" rm -f data/states.jsonl data/conceptual_.jsonl out//*.md")
	return 0


	if __name__ == "__main__":
	sys.exit(main())