File size: 6,683 Bytes
8b97eb8 | 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 | """Baseline LLM-as-agent solver for RealSR v3.
Runs the multi-turn equation-discovery agent on ONE public task and writes a
submission module (the agent's `<final_formula>`). In fixed-data mode, it can
optionally call the fixed-data numeric evaluator. In simulator/parallel mode,
evaluation is structure-only and is handled separately by
`harness/evaluate_parallel.py`. The agent sees ONLY the public task context; in
simulator mode it must collect observations through `<experiment>`.
Usage:
export OPENAI_API_KEY=... # or ANTHROPIC_API_KEY / OPENROUTER_API_KEY …
python run_baseline.py <task_dir> <model> [options]
<task_dir> path to a public task, e.g.
../tasks/typeI/cepheid_period_luminosity__M_W
<model> model alias (see call_llm_api.py: gpt5, gpt5mini,
claude-opus-4-7, gemini-3.1-pro, deepseek-reasoner, …)
Options:
--max-turns N agent turn budget (default 30)
--out DIR where to write <task_id>.py (default: ./submissions)
--simulator run in simulator-backed mode and enable <experiment>
--score fixed-data mode only: score it with the sibling numeric
harness (requires the private scoring/ tree to be present)
Batch all tasks:
for d in ../tasks/typeI/*/ ../tasks/typeII/*/ ; do
python run_baseline.py "$d" gpt5mini --out submissions
done
"""
from __future__ import annotations
import argparse
import json
import os
import sys
import time
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent))
from task import load_task # noqa: E402
from agent import conduct_exploration # noqa: E402
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("task_dir", help="path to a public task dir (tasks/<type>/<task>)")
ap.add_argument("model", help="model alias (see call_llm_api.py)")
ap.add_argument("--max-turns", type=int, default=30)
ap.add_argument("--out", default="submissions", help="output dir for <task_id>.py")
ap.add_argument("--traj-out", default=None,
help=("directory for per-turn trajectory checkpoints "
"(default: same as --out)"))
ap.add_argument("--simulator", nargs="?", const="simulator", default=None,
help=("enable simulator-backed mode. With current tasks, pass "
"`--simulator`; old named layouts may pass a simulator name."))
ap.add_argument("--score", action="store_true",
help="score the submission with the sibling harness (needs scoring/)")
ap.add_argument("--include-test-range", dest="include_test_range",
action="store_true", default=None,
help="include public metadata input train->test ranges in the task prompt (fixed real-data mode default)")
ap.add_argument("--no-include-test-range", dest="include_test_range",
action="store_false",
help="omit public metadata input train->test ranges from the task prompt")
args = ap.parse_args()
include_test_range = (
(args.simulator is None)
if args.include_test_range is None
else bool(args.include_test_range)
)
if args.simulator is not None:
include_test_range = False
task = load_task(
args.task_dir,
simulator=args.simulator,
show_test_range=include_test_range,
)
task_type = "typeII" if task.has_group_id else "typeI"
mode = "simulator" if args.simulator is not None else "fix"
objective = (
"objective=structure"
if args.simulator is not None
else f"metric={task.headline_metric}"
)
print(f"Task: {task.task_id} type={task_type} model={args.model} "
f"mode={mode} {objective} "
f"train_rows={len(task.train)}", flush=True)
out_dir = Path(args.out)
out_dir.mkdir(parents=True, exist_ok=True)
traj_dir = Path(args.traj_out) if args.traj_out else out_dir
traj_dir.mkdir(parents=True, exist_ok=True)
traj_path = traj_dir / f"{task.task_id}.traj.json"
def write_checkpoint(trial: dict) -> None:
payload = {
"meta": {
"task_id": task.task_id,
"task_dir": str(task.task_dir),
"task_type": task_type,
"mode": mode,
"model": args.model,
"max_turns": args.max_turns,
"include_test_range": include_test_range,
"checkpoint_path": str(traj_path),
"updated_at_unix": time.time(),
},
"trial": trial,
}
tmp_path = traj_path.with_suffix(traj_path.suffix + ".tmp")
with tmp_path.open("w") as fh:
json.dump(payload, fh, indent=2, sort_keys=True)
fh.write("\n")
os.replace(tmp_path, traj_path)
t0 = time.time()
initial_train_rows = len(task.train)
trial = conduct_exploration(task, model_name=args.model, max_turns=args.max_turns,
trial_info={"trial_id": f"{args.model}_{task.task_id}"},
checkpoint_fn=write_checkpoint)
eq = trial.get("submitted_equation") or ""
print(f"\n=== agent done ({time.time()-t0:.0f}s, status={trial.get('status')}, "
f"rounds={trial.get('rounds')}, tokens={trial.get('total_tokens')}, "
f"experiments={trial.get('n_experiments', 0)}, "
f"python_calls={trial.get('n_python_calls', 0)}, "
f"active_rows={max(0, len(task.train) - initial_train_rows)}) ===")
if not eq.strip():
print("agent produced no <final_formula>; nothing written.")
sys.exit(1)
out_path = out_dir / f"{task.task_id}.py"
out_path.write_text(eq)
print(f"submission written: {out_path}")
print(f"trajectory checkpoint: {traj_path}")
if args.score and args.simulator is not None:
raise SystemExit(
"--score is fixed-data only. For simulator/parallel runs, use "
"harness/evaluate_parallel.py to produce structure_score."
)
if args.score:
harness = Path(__file__).resolve().parent.parent / "harness"
sys.path.insert(0, str(harness.parent))
from harness import evaluate_on_test # noqa: PLC0415
res = evaluate_on_test(eq, task)
ns = res.get("numeric_score")
print(f"\nnumeric_score = {ns if ns is None else round(ns, 4)} "
f"(metric={res.get('metric')}, contract_ok={res.get('contract_ok')}, "
f"status={res.get('status')})")
if __name__ == "__main__":
main()
|