File size: 2,981 Bytes
53a1159 | 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 | """Frozen early/late/permanent freezing for the development-selected component."""
import argparse
import itertools
import json
from datetime import datetime, timezone
from pathlib import Path
import torch
from llm_memory_editability.bios_data import write_json
from llm_memory_editability.bios_direction import ROOT, digest
from llm_memory_editability.bios_direction_formation import edit_parent, train
ART = ROOT / "docs/development-artifacts/direction-causal-v1"
OUT = ROOT / "results/bios-direction-causal-v1"
PRED = ROOT / "docs/development-artifacts/direction-prediction-v1/predictor.json"
def freeze():
selected = json.loads(PRED.read_text())["selected_training_arm"].split("-")[0]
files = [
Path(__file__),
PRED,
ROOT / "src/llm_memory_editability/bios_direction_formation.py",
ROOT / "src/llm_memory_editability/bios_direction.py",
ROOT / "src/llm_memory_editability/bios_model.py",
ROOT / "docs/development-artifacts/direction-v1/selected.json",
]
payload = dict(
created=datetime.now(timezone.utc).isoformat(),
worlds=[4000, 4001, 4002, 4003],
seeds=[0, 1],
arms=[selected + "-" + name for name in ("freezeearly", "freezelate", "freezeall")],
methods=["func-soft-adam", "func-hard-gn"],
training_trajectories=24,
editing_trajectories=288,
hypothesis=(
"Early freezing and restoration should differ from late freezing "
"if early organization affects independent editability."
),
limitation="Development only; global impairment cannot establish selective necessity.",
files={str(p.relative_to(ROOT)): digest(p) for p in files},
)
if (ART / "lock.json").exists():
assert json.loads((ART / "lock.json").read_text())["files"] == payload["files"]
return
write_json(ART / "lock.json", payload)
(ART / "source.py").write_bytes(Path(__file__).read_bytes())
def run(worlds, device):
lock = json.loads((ART / "lock.json").read_text())
for p, h in lock["files"].items():
assert digest(ROOT / p) == h, p
for world, seed, arm in itertools.product(worlds, lock["seeds"], lock["arms"]):
out = OUT / f"world-{world}-seed-{seed}" / arm
train(world, seed, arm, device, out)
edit_parent(world, seed, arm, device, out, edit_arms=lock["methods"])
print(json.dumps(dict(world=world, seed=seed, arm=arm)), flush=True)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("command", choices=["freeze", "run"])
parser.add_argument("--worlds", nargs="+", type=int)
parser.add_argument("--device", default="cuda:0")
args = parser.parse_args()
torch.set_num_threads(2)
torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cudnn.allow_tf32 = False
if args.command == "freeze":
freeze()
else:
run(args.worlds, torch.device(args.device))
|