Research-Stack/.atlas/score.py

65 lines
2.1 KiB
Python
Executable file

#!/usr/bin/env python3
"""Generated by `atlas autoresearch`. Parse one metric out of a run's output
into the Atlas benchmark contract.
EDIT the METRIC line if the wrong number is being read, or rewrite parse() for
a richer evaluator (per-example feedback is what makes GEPA converge — see
.atlas/benchmark-contract or the atlas-optimize skill)."""
import argparse
import json
import os
import re
import sys
# autoresearch guessed this metric name from your --goal.
METRIC = "find_optimal_crossing" # <-- EDIT ME if this is the wrong metric
HIGHER_IS_BETTER = False
# Matches "<metric> = 0.74", "<metric>: 0.74", "<metric> 0.74".
PATTERN = re.compile(
re.escape(METRIC) + r"\s*[:=]?\s*([+-]?[0-9]*\.?[0-9]+(?:[eE][+-]?[0-9]+)?)"
)
def parse(text):
matches = PATTERN.findall(text)
if not matches:
return None
return float(matches[-1]) # last occurrence == final epoch / final eval
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--stdout", required=True, help="file holding the run's captured output")
args = ap.parse_args()
text = open(args.stdout, encoding="utf-8", errors="replace").read()
score = parse(text)
if score is None:
sys.stderr.write(
f"[score] could not find metric '{METRIC}' in the run output.\n"
" Fix: edit .atlas/score.py (METRIC=...), or pass an explicit\n"
" --benchmark to 'atlas autoresearch'.\n"
)
sys.exit(3)
direction = "higher is better" if HIGHER_IS_BETTER else "lower is better"
result = {
"score": score,
"examples": [{
"id": METRIC,
"score": score,
"pass": True,
"feedback": f"{METRIC} = {score} (parsed from run output; {direction})",
}],
"feedback": f"Final {METRIC} = {score}.",
}
out = os.environ.get("ATLAS_OPTIMIZE_RESULT")
payload = json.dumps(result)
if out:
with open(out, "w", encoding="utf-8") as fh:
fh.write(payload)
else:
sys.stdout.write(payload)
if __name__ == "__main__":
main()