benchmark/Belief/evaluation/README.md

Tie-aware Belief evaluation

This directory adds set-valued choice labels for benchmark rows with exactly zero position-derived confidence. It does not modify the released prompts, v1 label_choice, sample IDs, splits, profiles, retrieval contexts, or regression targets.

For a non-tied row (label_confidence > 0), choice scoring remains exact match against label_choice.

For an exact tie (label_confidence == 0), both equally held market outcomes are valid:

correct = normalize(predicted_choice) in normalize(valid_choices)

Invalid or unrelated outputs remain incorrect.

Files

• tied_valid_choices.parquet: set-valued labels for tied rows.
• tied_valid_choices.jsonl: the same metadata in JSONL format.
• tie_aware_scorer.py: reference scorer for saved predictions.
• build_tie_aware_metadata.py: reproducible metadata builder.
• validation_report.json: coverage and integrity checks.

The metadata contains:

sample_id, split, wallet, market_id, question,
label_choice_v1, label_confidence, valid_choices,
choice_status, outcome_source

Outcome names are recovered, in order of preference, from released Trade observations, Polymarket CLOB market metadata, other observed Belief stances for the same market, or an explicit complement map for standard binary labels such as yes/no and Democrat/Republican.

Example

python benchmark/Belief/evaluation/tie_aware_scorer.py predictions.parquet

The prediction file must include sample_id, label_choice, and predicted_choice (or specify another prediction column).

Tie-aware accuracy should be identified explicitly in result tables so it is not confused with the original v1 exact-match metric. No model inference needs to be rerun; saved predictions can be rescored.