reward: RLPR WebInstruct math grader ignores \boxed{...} answers, scoring correct answers 0

[philluiz2323]

Bug

_check_rlpr_webinstruct in src/trinity/orchestration/reward.py scores WebInstruct-verified RLPR items. For a math answer it feeds the raw candidate text straight into math_equal / normalize_math_answer (lines 484–487) without first extracting the answer — unlike _check_math, which calls extract_boxed / extract_last_number before comparing.

Since normalize_math_answer does not strip \boxed{...} (or pull an answer out of prose), a correct boxed answer never matches the bare gold value → scored 0.

Reproduce (offline)

from trinity.orchestration import reward as R
ref = {"ground_truth": "15", "source": "WebInstruct-verified-val_Avg2"}
R.score_text("rlpr", r"The area is \boxed{15}.", ref)   # -> 0.0  (should be 1.0)
R.score_text("rlpr", "The answer is 15", ref)           # -> 1.0  (plain number works)
R.score_text("math500", r"The area is \boxed{15}.", "15")  # -> 1.0 (math path extracts the box)

The worker is even instructed by format_hint("rlpr") to "use a boxed final answer for math items", so this is the expected output format — and it's graded wrong.

Impact

Every WebInstruct-verified RLPR math item answered with the requested \boxed{...} format is a false negative, understating accuracy on that benchmark and corrupting the reward signal for those items.

Distinct from

#116 / #122 (RLPR MMLU-Pro choice grading, letters E–J). This is the WebInstruct math path — a different branch and a different fix.

Fix

Extract the answer from the candidate (and gold) before the math compare, mirroring _check_math:

cand_math = extract_boxed(cand) or extract_last_number(cand) or cand
gold_math = extract_boxed(gold) or gold
if math_equal(cand_math, gold_math):
    return True
return normalize_math_answer(cand_math) == normalize_math_answer(gold_math)

Location

src/trinity/orchestration/reward.py, _check_rlpr_webinstruct (~lines 484–487).