solveedit_benchmark_2728 / code /question_eval_architecture.md
wenjieshu's picture
Add files using upload-large-folder tool
b8ea052 verified
|
Raw History Blame Contribute Delete
3.66 kB

Question-first evaluation v2

The formal evaluation unit is an audited atomic question, not a checker name.

QuestionSpec -> Evidence DAG -> EvidencePacket -> AnswerRecord -> role-aware score

Invariants

  • The original question, pass/partial/fail rubric, role, and weight are frozen in QuestionSpec.
  • A free-form decision_rule.type is never executable by resemblance. Only a registered normalizer may emit a deterministic PredicateSpec.
  • Every tool dependency is enforced by the DAG executor. A required failed dependency blocks downstream calls.
  • Detector evidence is scoped by case, check, image role, and resolved ROI.
  • An explicit detector no_detections result may establish zero; a missing row may not.
  • A detector bbox synthesized as a mask is insufficient for formal geometry.
  • Deterministic answers require a successful packet, a unique numeric band, and citable call IDs.
  • Penalty answers retain positive rubric-satisfaction semantics. Inversion happens once in the final scorer.
  • The independent image validity/quality gate is applied after question scoring; a failed gate forces the final case score to zero.
  • VLM fallback receives the exact question/rubric, compact tool facts, input/prediction, and question-specific overlays. It does not receive reference GT.
  • API keys are read from an environment variable at call time and are not CLI values or audit fields.

Current migration audit

The 271-case reviewed contract set compiles to 2,198 unique question specs:

  • 7 deterministic questions with explicitly normalized semantics;
  • 2,191 vlm_required questions;
  • 1,825 distinct source rule names preserved for migration analysis;
  • no missing resolved case and no duplicate procedure hash.

The deliberately small deterministic whitelist currently covers:

  • forbidden/non-exempt instance counts;
  • source-normalized zone clearance;
  • assigned vehicle slot containment with minimum area and color-evidence gating.

The accessible-vehicle rule remains vlm_required because containment alone cannot verify the accessibility symbol.

Entry points

Compile and audit without running models:

PYTHONPATH=src python scripts/run_question_eval_v2.py compile \
  --contracts-dir <contracts> --resolved-root <resolved-shards> --output-dir <output>

Replay legacy evidence without VLM calls:

PYTHONPATH=src python scripts/run_question_eval_v2.py replay \
  --contracts-dir <contracts> --resolved-root <resolved-shards> \
  --regions-tsv <regions> --detectors-tsv <fused-detections> --masks-tsv <masks> \
  --tool-calls-tsv <tool-calls> --output-dir <output>

Add --run-vlm --vlm-model <model> --resume to answer unresolved questions. Questions are batched once per case while retaining separate packets and answer records. Use --num-shards and --shard-index for parallel workers.

Merge completed shards with conflict and provenance checks:

python scripts/merge_question_eval_v2_shards.py \
  --shard-dir <shard-0> --shard-dir <shard-1> \
  --question-specs <shard-0/question_specs.jsonl> --output-dir <merged>

Golden regressions

  • Black spiked crystals: 0 = pass, 1–2 = partial, 3+ = fail.
  • Cleared stairway: 0 output vehicles = pass; presence semantics cannot be reversed.
  • Parking slot: containment [0.65, 0.85) = partial, rather than multiplying 0.85 * 0.5.
  • Required DAG dependencies block downstream tools.
  • Bbox fallback cannot become a formal mask.
  • Formal answers must cite tool calls.
  • Runtime NumPy masks are persisted as compact summaries, not raw pixel arrays.
  • Multiple unresolved questions in one case use one batch VLM request.