SlashYear-Bench
1,000 questions about dated history, asked closed-book: the model gets the question and nothing else. Every answer key comes from the same sourced record as the rest of the site. Event sentences are quoted from a numbered Wikipedia revision with the date removed. Leaders are taken from the checked leader lists, No. 1 singles from the Hot 100 lists and dollar values from the BLS consumer price index. Each item carries the source it was built from.
Questions and answers: /api/bench/items.jsonl · scores: /api/bench/summary.json · code: bench/. Run 2026-10-27.
| Model | Overall | Year of an event | Billboard No. 1 on a chart date | Who held an office in a year | What $100 was worth in another year |
|---|---|---|---|---|---|
| gpt-6.1-sol | 96.7% | 93.3% | 98.0% | 99.2% | 100.0% |
| gpt-6-astra | 96.5% | 93.0% | 98.0% | 98.8% | 100.0% |
| gpt-6-luna | 90.7% | 89.8% | 78.0% | 97.6% | 98.7% |
The questions
- Year of an event (400): a dated sentence with its date removed. Sentences that give away the year, or that appear under more than one year, are left out. Scored on the exact year. Also reported within one year.
- Billboard No. 1 (200): the song at No. 1 on a chart issue date, 1958 on. Scored on the title.
- Office holders (250): only state, post and year combinations with exactly one recorded holder, so there is one right answer.
- Dollar values (150): $100 in one year in another year’s dollars, by the annual CPI-U. Correct within 3%.