File size: 9,006 Bytes
c10636a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 | #!/usr/bin/env python3
"""Run a recoverable evaluator-only Agent rollout from a public test release.
This wrapper deliberately reuses the durable engineering-retry collector used
for teacher rollout. It differs in two important ways:
* the input must be a ``development_subset`` release marked
``training_prohibited=true``;
* it never runs SFT eligibility, reward scoring, or training-data staging.
Every successful trace is retained and linked into ``rollouts/test/merged``.
Only case IDs without a terminal success are retried. ``agent-results.jsonl``
is the normalized, one-row-per-case input for the post-hoc private-gold judge.
"""
from __future__ import annotations
import argparse
import json
import sys
from collections.abc import Mapping
from pathlib import Path
from typing import Any
REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from scripts.trajectory.run_teacher_rollout_autopilot import (
_read_json,
_read_jsonl,
_run_engineering_retries,
_write_json,
_write_jsonl,
)
from eviscene.eval.public_release import load_public_release
from eviscene.eval.run_artifacts import load_jsonl_objects
SCHEMA_VERSION = "ifv-agent-test-rollout-v1"
VALID_VERDICTS = frozenset({"real", "fake"})
def _read_release_case_ids(benchmark: Path) -> list[str]:
rows = load_jsonl_objects(benchmark)
case_ids = [str(row.get("case_id") or "").strip() for row in rows]
if not case_ids or any(not case_id for case_id in case_ids):
raise ValueError("test release contains an empty case_id")
if len(case_ids) != len(set(case_ids)):
raise ValueError("test release contains duplicate case_id values")
return case_ids
def _test_release_is_training_prohibited(release_root: Path) -> None:
manifest_path = release_root / "manifest.json"
manifest = _read_json(manifest_path)
if manifest.get("release_stage") != "development_subset":
raise ValueError(
"test rollout only accepts release_stage=development_subset; "
"refusing a teacher/train release"
)
if manifest.get("training_prohibited") is not True:
raise ValueError(
"test rollout release must set training_prohibited=true"
)
def _build_agent_results(
*,
merged_dir: Path,
target_case_ids: list[str],
) -> list[dict[str, Any]]:
provenance_path = merged_dir / "trace-provenance.jsonl"
provenance_rows = _read_jsonl(provenance_path)
by_case: dict[str, Mapping[str, Any]] = {}
for row in provenance_rows:
case_id = str(row.get("case_id") or "").strip()
if not case_id:
raise ValueError(f"trace provenance row lacks case_id: {provenance_path}")
if case_id in by_case:
raise ValueError(f"duplicate merged trace provenance: {case_id}")
by_case[case_id] = row
target = set(target_case_ids)
unexpected = sorted(set(by_case) - target)
if unexpected:
raise ValueError(f"merged trace has unexpected case ID: {unexpected[0]}")
results: list[dict[str, Any]] = []
for case_id in target_case_ids:
provenance = by_case.get(case_id)
if provenance is None:
results.append(
{
"case_id": case_id,
"status": "error",
"error": "engineering attempts exhausted without terminal success",
"training_prohibited": True,
}
)
continue
trace_name = Path(str(provenance.get("source_trace") or "")).name
trace_path = merged_dir / "traces" / trace_name
if not trace_path.is_file():
raise FileNotFoundError(f"merged trace is missing: {trace_path}")
trace = _read_json(trace_path)
verdict = str(trace.get("verdict") or "").lower()
termination = str(trace.get("termination") or "")
if termination != "success" or verdict not in VALID_VERDICTS:
raise ValueError(
"merged trace is not terminal success: "
f"{case_id} termination={termination!r} verdict={verdict!r}"
)
trace_case_id = str(trace.get("case_id") or "").strip()
if trace_case_id != case_id:
raise ValueError(
f"merged trace case mismatch: result={case_id}, trace={trace_case_id}"
)
results.append(
{
"case_id": case_id,
"status": "success",
"trace_path": str(trace_path.relative_to(merged_dir).as_posix()),
"verdict": verdict,
"confidence": trace.get("confidence"),
"termination": termination,
"source_attempt": provenance.get("source_attempt"),
"source_trace": provenance.get("source_trace"),
"training_prohibited": True,
}
)
return results
def run_agent_test_rollout(
*,
benchmark: Path,
output_dir: Path,
profile: str,
concurrency: int,
base_seed: int,
timeout: float,
maximum_attempts: int,
) -> dict[str, Any]:
"""Run test cases with durable engineering recovery only."""
benchmark = benchmark.expanduser().resolve()
output_dir = output_dir.expanduser().resolve()
release = load_public_release(benchmark)
_test_release_is_training_prohibited(release.root)
target_case_ids = _read_release_case_ids(benchmark)
if output_dir.exists() and any(output_dir.iterdir()):
raise FileExistsError(f"output directory must be new or empty: {output_dir}")
if concurrency < 1:
raise ValueError("concurrency must be positive")
if maximum_attempts < 1:
raise ValueError("maximum_attempts must be positive")
if timeout <= 0:
raise ValueError("timeout must be positive")
output_dir.mkdir(parents=True, exist_ok=False)
_write_json(
output_dir / "run-config.json",
{
"schema_version": SCHEMA_VERSION,
"benchmark": str(benchmark),
"release_id": release.release_id,
"release_stage": release.release_stage,
"training_prohibited": True,
"profile": profile,
"concurrency": concurrency,
"base_seed": base_seed,
"timeout": timeout,
"maximum_attempts": maximum_attempts,
"target_case_count": len(target_case_ids),
},
)
(output_dir / "target-case-list.txt").write_text(
"".join(f"{case_id}\n" for case_id in target_case_ids),
encoding="utf-8",
)
merged_dir, merged_manifest = _run_engineering_retries(
group_name="test",
benchmark=benchmark,
pipeline_dir=output_dir,
target_ids=target_case_ids,
profile=profile,
rollout_concurrency=concurrency,
base_seed=base_seed,
timeout=timeout,
maximum_attempts=maximum_attempts,
candidates_per_case=1,
)
results = _build_agent_results(
merged_dir=merged_dir,
target_case_ids=target_case_ids,
)
_write_jsonl(output_dir / "agent-results.jsonl", results)
result = {
"schema_version": SCHEMA_VERSION,
"benchmark": str(benchmark),
"release_id": release.release_id,
"training_prohibited": True,
"target_case_count": len(target_case_ids),
"successful_case_count": sum(
row.get("status") == "success" for row in results
),
"engineering_error_count": sum(
row.get("status") != "success" for row in results
),
"merged_dir": str(merged_dir),
"merged_manifest": str(merged_dir / "run_manifest.json"),
"agent_results": str(output_dir / "agent-results.jsonl"),
"retry_manifest": merged_manifest,
}
_write_json(output_dir / "summary.json", result)
return result
def main() -> int:
parser = argparse.ArgumentParser(
description="Run an isolated Agent test release with engineering-only retries."
)
parser.add_argument("--benchmark", required=True)
parser.add_argument("--output-dir", required=True)
parser.add_argument("--profile", default="teacher-gemini")
parser.add_argument("--concurrency", type=int, default=10)
parser.add_argument("--base-seed", type=int, default=1729)
parser.add_argument("--timeout", type=float, default=1800.0)
parser.add_argument("--maximum-attempts", type=int, default=4)
args = parser.parse_args()
result = run_agent_test_rollout(
benchmark=Path(args.benchmark),
output_dir=Path(args.output_dir),
profile=args.profile,
concurrency=args.concurrency,
base_seed=args.base_seed,
timeout=args.timeout,
maximum_attempts=args.maximum_attempts,
)
print(json.dumps(result, ensure_ascii=False, indent=2))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|