Skip to content

Cache judge responses for reruns #353

Description

@ScuttleBot

Priority 5: Cache Judge Responses (5-10% speedup on reruns)

Problem

When rerunning benchmarks for regression testing or model comparisons, we re-grade identical transcripts.

Solution

Cache judge responses keyed by (task_id, transcript_hash).

Implementation

import hashlib
from pathlib import Path

JUDGE_CACHE_DIR = Path(".judge_cache")

def cache_key(task_id: str, transcript: list) -> str:
    transcript_str = json.dumps(transcript, sort_keys=True)
    hash_hex = hashlib.sha256(transcript_str.encode()).hexdigest()[:16]
    return f"{task_id}:{hash_hex}"

def get_cached_grade(task_id: str, transcript: list) -> dict | None:
    key = cache_key(task_id, transcript)
    cache_path = JUDGE_CACHE_DIR / f"{key}.json"
    if cache_path.exists():
        return json.loads(cache_path.read_text())
    return None

Expected Impact

  • Minimal for first run
  • Significant for reruns (regression testing)
  • Reduces judge API costs over time

Caveats

  • Only cache deterministic transcripts
  • May need opt-in flag (--use-judge-cache)
  • Cache invalidation when rubric changes

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions