File size: 3,779 Bytes
22decf5 | 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 | """Command line entry point: `ptb setup | run | eval`."""
import argparse
import json
import time
from pathlib import Path
from . import config
from .evaluate import evaluate
from .provision import setup
from .solve import solve
def run(args: argparse.Namespace) -> None:
"""Full benchmark run: agent phase, then evaluation of its final model."""
safe = str.maketrans("/:[]", "____") # same run dir naming as upstream, so its scripts/ keep working
run_dir = (config.RESULTS_DIR
/ f"{args.agent}_{args.agent_model.translate(safe)}_{args.hours}h{args.experiment_name}"
/ f"{args.task}_{args.model.translate(safe)}_{time.strftime('%Y%m%d-%H%M%S')}")
run_dir.mkdir(parents=True)
print(f"Run directory: {run_dir}")
solve(run_dir, args.task, args.model, args.agent, args.agent_model, args.hours, args.container)
if not (run_dir / "final_model").is_dir():
raise SystemExit("The agent produced no final_model; nothing to evaluate.")
report(evaluate(run_dir, args.task, str(run_dir / "final_model")))
def report(metrics: dict | None) -> None:
if metrics is None:
raise SystemExit("Evaluation failed; see the final_eval_*.txt logs.")
print(json.dumps(metrics, indent=2))
def main() -> None:
parser = argparse.ArgumentParser(prog="ptb", description=__doc__)
commands = parser.add_subparsers(dest="command", required=True)
p = commands.add_parser("setup", help="build the container images, download models and test data")
p.add_argument("--containers", nargs="*", default=[config.AGENT_CONTAINER, config.EVAL_CONTAINER],
choices=config.DEFINITIONS, help="upstream container definitions to build")
p.add_argument("--full-cache", action="store_true",
help="prefill the HF cache with upstream's 14 models and 382 dataset splits")
p.add_argument("--models", nargs="*", default=[], help="HF ids of base models to download")
p.add_argument("--tasks", nargs="*", default=[], choices=config.TASKS, help="benchmarks to fetch test data for")
p.set_defaults(func=lambda a: setup(a.containers, a.models, a.tasks, a.full_cache))
p = commands.add_parser("run", help="let an agent post-train a model, then evaluate the result")
p.add_argument("--task", required=True, choices=config.TASKS)
p.add_argument("--model", required=True, help="HF id of the base model to post-train")
p.add_argument("--agent", default="claude_non_api", choices=config.AGENTS, help="upstream agent scaffold")
p.add_argument("--agent-model", required=True, help="API model driving the agent, e.g. claude-opus-5")
p.add_argument("--container", default=config.AGENT_CONTAINER, choices=config.DEFINITIONS,
help="container for the agent phase; upstream picks one per cell")
p.add_argument("--hours", type=int, default=10, help="time budget of the agent")
# Upstream's EXPERIMENT_NAME: a suffix on the results dir, so the repeated seeds of one cell
# land in the `..._10h_run1/_run2/_run3` directories `scripts/collect.py` aggregates over.
p.add_argument("--experiment-name", default="", help="suffix for the results directory, e.g. _run1")
p.set_defaults(func=run)
p = commands.add_parser("eval", help="evaluate a model only (e.g. a base model as baseline)")
p.add_argument("--task", required=True, choices=config.TASKS)
p.add_argument("--model", required=True, help="model directory or HF id")
p.add_argument("--out-dir", required=True, type=Path)
p.add_argument("--limit", type=int, default=-1, help="number of samples; -1 = full benchmark")
p.set_defaults(func=lambda a: report(evaluate(a.out_dir, a.task, a.model, a.limit)))
args = parser.parse_args()
args.func(args)
|