File size: 6,283 Bytes
6cd33cb | 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 | #!/usr/bin/env python3
"""Run the locked RAID-only pairwise-CE neologism training recipe."""
from __future__ import annotations
import argparse
import os
import subprocess
import sys
from dataclasses import dataclass
from pathlib import Path
@dataclass(frozen=True)
class ModelProfile:
model: str
ai_token: str
human_token: str
pilot_lr: float
continuation_steps: int
continuation_lr: float | None = None
continuation_pairs: int | None = None
schedule_total_steps: int | None = None
PROFILES = {
"gemma": ModelProfile(
model="google/gemma-4-E4B-it",
ai_token="<ai>",
human_token="<human>",
pilot_lr=1e-3,
continuation_steps=75,
continuation_lr=1e-4,
continuation_pairs=5_000,
schedule_total_steps=625,
),
"llama": ModelProfile(
model="meta-llama/Llama-3.1-8B-Instruct",
ai_token="<AIGEN>",
human_token="<REAL>",
pilot_lr=1e-4,
continuation_steps=0,
),
}
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--model-family", choices=PROFILES, required=True)
parser.add_argument("--dataset-disk", type=Path, required=True)
parser.add_argument("--ai-init", type=Path, required=True)
parser.add_argument("--human-init", type=Path, required=True)
parser.add_argument("--output-dir", type=Path, required=True)
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--model", help="Override the Hugging Face model ID/path.")
parser.add_argument("--runner", type=Path)
parser.add_argument("--beemo-pairs", type=int, default=2_163)
parser.add_argument("--bootstrap-resamples", type=int, default=1_000)
parser.add_argument("--dry-run", action="store_true")
return parser.parse_args()
def run(command: list[str], *, repo_root: Path, dry_run: bool) -> None:
printable = " ".join(subprocess.list2cmdline([arg]) for arg in command)
print(f"+ {printable}", flush=True)
if dry_run:
return
env = os.environ.copy()
env["PYTHONPATH"] = os.pathsep.join(
part
for part in (str(repo_root), env.get("PYTHONPATH", ""))
if part
)
subprocess.run(command, cwd=repo_root, env=env, check=True)
def common_command(
*,
python: str,
runner: Path,
args: argparse.Namespace,
profile: ModelProfile,
output_dir: Path,
ai_init: Path,
human_init: Path,
) -> list[str]:
return [
python,
"-u",
str(runner),
"--dataset-disk",
str(args.dataset_disk),
"--train-split",
"standard_train_expanded6",
"--test-split",
"standard_test",
"--output-dir",
str(output_dir),
"--model",
args.model or profile.model,
"--ai-token",
profile.ai_token,
"--human-token",
profile.human_token,
"--prompt-template",
"Write {token} text.",
"--ai-init",
str(ai_init),
"--human-init",
str(human_init),
"--objective",
"ai_pairwise",
"--max-length",
"512",
"--batch-size",
"8",
"--eval-batch-size",
"8",
"--min-lr",
"1e-5",
"--warmup-steps",
"20",
"--beta",
"1",
"--bootstrap-resamples",
str(args.bootstrap_resamples),
"--seed",
str(args.seed),
]
def main() -> None:
args = parse_args()
profile = PROFILES[args.model_family]
repo_root = Path(__file__).resolve().parents[1]
runner = args.runner or repo_root / "scripts/run_gemma_expanded_pairwise_ce.py"
output_dir = args.output_dir.resolve()
train_dir = output_dir / "train"
eval_dir = output_dir / "eval"
if output_dir.exists() and any(output_dir.iterdir()):
raise FileExistsError(f"Refusing to overwrite non-empty {output_dir}")
train_dir.mkdir(parents=True, exist_ok=True)
train = common_command(
python=sys.executable,
runner=runner,
args=args,
profile=profile,
output_dir=train_dir,
ai_init=args.ai_init,
human_init=args.human_init,
)
train.extend(
[
"--pilot-pairs",
"500",
"--pilot-epochs",
"2",
"--pilot-lrs",
str(profile.pilot_lr),
"--pilot-beemo-pairs",
"250",
]
)
if args.model_family == "gemma":
assert profile.continuation_lr is not None
assert profile.continuation_pairs is not None
assert profile.schedule_total_steps is not None
train.extend(
[
"--full-pairs",
str(profile.continuation_pairs),
"--full-epochs",
"1",
"--full-lr",
str(profile.continuation_lr),
"--stop-after-steps",
str(profile.continuation_steps),
"--schedule-total-steps",
str(profile.schedule_total_steps),
"--eval-every",
str(profile.continuation_steps),
"--beemo-pairs",
str(args.beemo_pairs),
]
)
run(train, repo_root=repo_root, dry_run=args.dry_run)
return
train.append("--pilot-only")
run(train, repo_root=repo_root, dry_run=args.dry_run)
# Llama continuation consistently regressed the fixed monitor, so evaluate
# the 125-update pilot directly on both authoritative test populations.
selected_ai = train_dir / "pilot/lr_0.0001/final/tokens/ai_token.pt"
eval_dir.mkdir(parents=True, exist_ok=True)
evaluate = common_command(
python=sys.executable,
runner=runner,
args=args,
profile=profile,
output_dir=eval_dir,
ai_init=selected_ai,
human_init=args.human_init,
)
evaluate.extend(
[
"--eval-only",
"--eval-targets",
"beemo,raid",
"--beemo-pairs",
str(args.beemo_pairs),
]
)
run(evaluate, repo_root=repo_root, dry_run=args.dry_run)
if __name__ == "__main__":
main()
|