Fix scoring bugs in 11 evaluation scripts
Eleven evaluation scripts in evaluation_scripts/2025_10_23/ score some answers differently from what the task text, the script's own comments, or its evaluation notes specify. Each fix is a local change of a few lines. All eleven scripts are in the test set (none is in the public dev set). None of them is touched by #6 or #8; #9 also changes county_AQI.py, on other lines, and the two versions merge cleanly.
Changes
| Script | Problem | Effect of the fix |
|---|---|---|
baltimore_marc_station.py |
The answer is padded to, and cut off at, 13 station slots, one per key of EXPECTED_STATIONS, but five of the keys are alternate names of two stations ("Laurel Race Track"/"Laurel Park"; "Riverdale"/"Riverdale Park"/"Riverdale Park Town Center"). The ground truth has 10 intermediate stops, so three empty slots always fail. |
Answers are padded to, and cut off at, 10 slots. A complete answer can score 1.0 instead of at most 10/13 ≈ 0.77, and an answer that lists more than 10 stations is evaluated on its first 10 instead of its first 13. |
gas_stations.py |
The task requires the last station to be "within 500 miles of Los Angeles", and EVAL_NOTES relaxes the 300–500-mile rule for consecutive stations to 290–510. The final leg is nevertheless checked against 290–510, so a last station closer than 290 miles to Los Angeles fails a critical check. | The final leg is checked against the 510-mile maximum only. |
ingredient_safety.py |
Brands are matched by a SequenceMatcher ratio of at least 0.8, which rejects "Jo Malone" for "Jo Malone London" (0.727) and "Christian Dior" for "Dior" (0.471). A miss fails a critical node and zeroes that brand's third of the task. |
A name below the threshold also matches when it is the required name with words added before or after it ("Christian Dior", "Dior Parfums", "Chanel Paris") or the first two or more words of the required name ("Jo Malone"). Such a match scores at the threshold, so a closer spelling in the same answer is still preferred. Names that matched before give the same result, and fragments such as "Jo" or "London" still do not match. |
ikea_shopping_list.py |
extract_numeric_price returns 0.0 when the price string has extra text ("$179.00 USD", "$29.99 each"), so an item whose price the judge verified adds $0 to the total that the critical budget node checks. |
A string with words around a single amount yields that amount, preferring the number after a dollar sign ("2 for $10" gives 10). Strings that parsed before give the same value. A string with several amounts ("Was $249.00, now $199.00") still gives 0.0, since which one is the price cannot be told. The budget verdict can change in either direction. |
county_AQI.py |
The GDP-year check builds its claim from population_year, and its description says "population data". |
The check uses gdp_year. Its verdict changes when exactly one of the two years is inside the window. |
active_stock.py |
The two tolerance notes in the judge instructions do not fit their requirements. The market-cap note reads (+ 0.2 billion), such as "15.1 million", a unit left over from the trading-volume note. The trading-volume requirement is a minimum ("over 10 million"), but its note reads (+ 0.2 million), such as "10.1 million", a value that already meets it, as if copied from the market-cap note, whose requirement is a maximum. |
The notes read such as "15.1 billion" and (- 0.2 million), such as "9.9 million". They should affect only market caps between $15.0B and $15.2B and volumes between 9.8 and 10 million shares, which the judge is now told to tolerate. |
best_selling_foundations.py |
The concealer and highlighter claims interpolate the product name without a guard, so a missing name reaches the judge as "a concealer product 'None'". The foundation claim guards the same case. | The quoted name is left out of the claim when it is missing. |
drone_delivery.py |
When the answer gives no country of origin for the first company, the judge receives "'None' and '<country>' are different countries". The first company's existence check is in a sibling branch, so it does not prevent the call. | The check fails without a judge call when either country is missing. |
rl_course_book.py |
add_parallel(..., parent_node=parent_node) misspells the parent keyword. Pydantic ignores the unknown keyword, so each university's instructor subtree is attached to the root instead of the course node. |
The subtree moves to the course node, where the call site puts it. Scores become stricter; see the next section, which also describes an alternative. The comment on the subtree's critical=False now states what that flag does. |
popular_ai_books.py |
list(set(titles))[:5] orders titles by string hash, which Python randomizes per process, so the five books that are evaluated vary between runs when the answer lists more than five. |
Duplicates are removed in order of appearance, so the first five titles are evaluated in every run. |
paper_collaboration.py |
info.first_author.lower() raises AttributeError when the answer gives no first author (the guard above it is commented out), and the answer gets no result. |
A missing first author becomes an empty string, as the script already does for arxiv_link and advisor_name. The answer is scored (0) instead of raising. |
rl_course_book: the fix makes scores stricter
Within each university, the course node is SEQUENTIAL (the course page exists, the book is recommended on it, the instructors are correct), and the university node is SEQUENTIAL over the course node and the book node. With the typo, the root averages six children behind the uniqueness gate: three university nodes and three detached instructor nodes. Instructor names therefore carry half of the task score, and an instructor check is sent to the judge even when that university's course page failed.
With the fix, the root averages the three university nodes. A wrong instructor lowers the course node to 0.5 rather than 0, as its critical=False intends, but the SEQUENTIAL university node then skips that university's book checks. The released comment on that flag ("instructor error shouldn't fail entire university") did not anticipate this, so the comment now describes only what the flag does.
Decision point. The instructor subtree could instead become the last child of the university node. A wrong instructor then lowers only its own share, but a failed book check skips the instructor check, which now comes after it. Scores with the fake judge rejecting the claims named in each row:
| Answer | Released | This PR | Alternative: instructors last |
|---|---|---|---|
| Fully correct | 1.000 | 1.000 | 1.000 |
| Correct except one university's instructors | 0.833 | 0.750 | 0.889 |
| Correct except every university's instructors | 0.500 | 0.250 | 0.667 |
| One university's course page fails, its instructor check passes | 0.833 | 0.667 | 0.667 |
| Correct except every university's purchase link | 0.750 | 0.500 | 0.333 |
This PR keeps the subtree where the call site puts it, since moving it is a rubric change rather than a typo fix. It can switch to the alternative if preferred.
Verification
The offline harness proposed in OSU-NLP-Group/Mind2Web-2#12 runs a script with a fake judge and a synthetic web cache under four synthetic answer policies. The eleven scripts were run under all four policies before and after the fixes, with no crashes, and the rubric trees changed only where expected:
baltimore_marc_station: three station slots disappear under every policy. Under three policies they are empty slots that always failed. Under the 12-station policy they are the 11th and 12th listed stations, which passed, and one empty slot, so that run's score goes from 0.923 to 1.0.rl_course_book: the instructor subtrees move under the course nodes (all four policies).drone_delivery, empty answer: the different-country check isfailedinstead ofskipped(score 0 either way).active_stock, hash policy: one verdict flips, because that fake judge derives its verdict from the prompt text, which the fix changes.
The synthetic policies do not reach most of the situations fixed here, so each fix was also run on a targeted input with the released script and the fixed script. The fake judge accepts every claim unless stated otherwise.
| Script | Input | Released | Fixed |
|---|---|---|---|
gas_stations |
Five 400-mile legs; the last station is 182 miles from Los Angeles | 0.8 | 1.0 |
drone_delivery |
The first company has no country of origin | 0.5; two judge requests contain 'None' and |
0.0; none do |
best_selling_foundations |
Concealer and highlighter without product names | Eight judge requests contain product 'None' |
None do |
paper_collaboration |
No first author | AttributeError |
Scored, 0.0 |
rl_course_book |
The five answers in the table above; the fake judge rejects the claims named there | See above | See above |
popular_ai_books |
Six titles, PYTHONHASHSEED set to 1, 2, and 3 |
A different title is dropped for each seed | The sixth title is dropped every time |
Direct calls of the two patched helper functions:
| Function | Input | Released | Fixed |
|---|---|---|---|
ingredient_safety.calculate_brand_similarity |
"Jo Malone" / "Jo Malone London" | 0.727 | 0.800 |
| "Christian Dior" / "Dior" | 0.471 | 0.800 | |
| "Dior Parfums" / "Dior" | 0.533 | 0.800 | |
| "Chanel Paris" / "Chanel" | 0.706 | 0.800 | |
| "JO MALONE LONDON" / "Jo Malone London", "Chanell" / "Chanel" | 1.000, 0.923 | unchanged | |
| "Jo", "London", "Malone" / "Jo Malone London" | 0.250, 0.600, 0.600 | unchanged | |
| "Jo Loves" / "Jo Malone London" | 0.476 | 0.476 | |
| "Chanel" / "Dior", "Tom Ford" / "Chanel" | 0.000 | 0.000 | |
ingredient_safety.find_matching_fragrance for "Jo Malone London" |
Fragrances listed as "Jo Malone", then "Jo Malone London" | the second | the second |
| Only "Jo Malone" | none | "Jo Malone" | |
ikea_shopping_list.extract_numeric_price |
"$179.00 USD", "$29.99 each", "From $49", "179 USD" | 0.00, 0.00, 0.00, 0.00 | 179.00, 29.99, 49.00, 179.00 |
| "2 for $10" | 0.00 | 10.00 | |
| "$179.00", "$1,299.00", "179" | 179.00, 1299.00, 179.00 | unchanged | |
"Was $249.00, now $199.00", "Price unavailable", "", None |
0.00 | 0.00 |
Scores and versioning
These fixes change the scores of affected answers on these eleven tasks, so results computed with the released scripts and with the fixed scripts are not directly comparable there. The fixes are applied in place in 2025_10_23, like #6, and the previous dataset revision remains available for reproducing earlier numbers; if they should instead ship as a new eval-version directory, the same eleven files can be placed there. A changelog entry should accompany the merge, for example: "Fixed scoring bugs in 11 evaluation scripts (active_stock, baltimore_marc_station, best_selling_foundations, county_AQI, drone_delivery, gas_stations, ikea_shopping_list, ingredient_safety, paper_collaboration, popular_ai_books, rl_course_book); scores of affected answers change."
Not changed
- Rubric design choices, such as SEQUENTIAL cascades, first-N conventions, and missing distinctness checks, are left as written.
- Several scripts compute their reference dates from the evaluation clock (
datetime.now()); #9 lets an evaluation set that date.
🤖 Generated with Claude Code