id large_stringlengths 14 31 | family large_stringclasses 58
values | domain large_stringclasses 17
values | category large_stringclasses 2
values | difficulty int64 1 8 | prompt_style large_stringclasses 4
values | variant large_stringclasses 4
values | pair_id large_stringlengths 20 33 ⌀ | split large_stringclasses 1
value | template_group large_stringlengths 19 38 | prompt large_stringlengths 185 13.8k | answer large_stringlengths 1 181 | answer_type large_stringclasses 10
values | answer_part_types listlengths 2 2 ⌀ | grading_rule large_stringclasses 10
values | answer_format large_stringclasses 10
values | problem_hash large_stringlengths 20 20 | prompt_chars int64 185 13.8k | metadata large_stringlengths 35 11.9k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
api_pagination-000001 | api_pagination | agentic_tool_use | applied | 1 | plain | base | null | train | api_pagination/plain/p0/s0 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 4 | int | null | exact integer match | an integer (digits only, no commas or units) | b44072eb48443c8988c7 | 1,477 | {"log":[{"cursor":"cur_bc1d0","items":[{"id":"inv-3635","status":"void","total":397},{"id":"inv-3220","status":"paid","total":510},{"id":"inv-7954","status":"void","total":10},{"id":"inv-2710","status":"open","total":108}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_986d8","items":[{"id":"inv-7070","... |
api_pagination-000002 | api_pagination | agentic_tool_use | applied | 1 | plain | base | null | train | api_pagination/plain/p0/s1 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 1 | int | null | exact integer match | an integer (digits only, no commas or units) | 865228424586f1502e5b | 1,426 | {"log":[{"cursor":"cur_2258d","items":[{"id":"inv-6432","status":"paid","total":101}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_42261","items":[{"id":"inv-4632","status":"open","total":743},{"id":"inv-9727","status":"void","total":11},{"id":"inv-9385","status":"open","total":614}],"next":"cur_5e130... |
api_pagination-000003 | api_pagination | agentic_tool_use | applied | 1 | plain | base | null | train | api_pagination/plain/p0/s2 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 4 | int | null | exact integer match | an integer (digits only, no commas or units) | 9a2dbad3e8d9f15dbc8b | 1,476 | {"log":[{"cursor":"cur_45cc3","items":[{"id":"inv-7279","status":"open","total":205},{"id":"inv-8227","status":"open","total":53}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_67ce0","items":[{"id":"inv-7616","status":"paid","total":163},{"id":"inv-2338","status":"open","total":768},{"id":"inv-8891","... |
api_pagination-000004 | api_pagination | agentic_tool_use | applied | 1 | plain | distractor | null | train | api_pagination/plain/p0/s3 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 6 | int | null | exact integer match | an integer (digits only, no commas or units) | 332ab81c1d87119c090f | 1,625 | {"log":[{"cursor":"cur_8c1b4","items":[{"id":"inv-8031","status":"paid","total":227}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_3cbc0","items":[{"id":"inv-5914","status":"void","total":773},{"id":"inv-9368","status":"paid","total":535},{"id":"inv-7910","status":"paid","total":147}],"next":null,"que... |
api_pagination-000007 | api_pagination | agentic_tool_use | applied | 2 | plain | base | null | train | api_pagination/plain/p0/s0 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 4 | int | null | exact integer match | an integer (digits only, no commas or units) | 838c2e94210c6196d3ec | 1,831 | {"log":[{"cursor":"cur_36381","items":[{"id":"inv-5730","status":"paid","total":879},{"id":"inv-1507","status":"open","total":610},{"id":"inv-7237","status":"paid","total":126},{"id":"inv-4878","status":"paid","total":839}],"next":"cur_22d1f","query":"status=any&customer=K-17"},{"cursor":"cur_dac91","items":[{"id":"inv... |
api_pagination-000008 | api_pagination | agentic_tool_use | applied | 2 | plain | base | null | train | api_pagination/plain/p0/s1 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 5 | int | null | exact integer match | an integer (digits only, no commas or units) | c200936b675abc522d34 | 1,779 | {"log":[{"cursor":"cur_56faa","items":[{"id":"inv-1129","status":"paid","total":348}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_2e1b0","items":[{"id":"inv-5084","status":"paid","total":946},{"id":"inv-6145","status":"paid","total":175}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur... |
api_pagination-000009 | api_pagination | agentic_tool_use | applied | 2 | plain | counterfactual_base | pair-api_pagination-000009 | train | api_pagination/plain/p0/s2 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 2 | int | null | exact integer match | an integer (digits only, no commas or units) | 98bfbd3fee9bdaa832f9 | 1,778 | {"log":[{"cursor":"cur_b527e","items":[{"id":"inv-7942","status":"paid","total":222}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_5a256","items":[{"id":"inv-9331","status":"open","total":139},{"id":"inv-5623","status":"open","total":510},{"id":"inv-5538","status":"void","total":99},{"id":"inv-7985","... |
api_pagination-000009-cf | api_pagination | agentic_tool_use | applied | 2 | plain | counterfactual | pair-api_pagination-000009 | train | api_pagination/plain/p0/s2 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 7 | int | null | exact integer match | an integer (digits only, no commas or units) | 9d8d413262edfd674506 | 1,778 | {"log":[{"cursor":"cur_b527e","items":[{"id":"inv-7942","status":"paid","total":222}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_5a256","items":[{"id":"inv-9331","status":"open","total":139},{"id":"inv-5623","status":"open","total":510},{"id":"inv-5538","status":"void","total":99},{"id":"inv-7985","... |
api_pagination-000010 | api_pagination | agentic_tool_use | applied | 2 | plain | base | null | train | api_pagination/plain/p0/s3 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 2 | int | null | exact integer match | an integer (digits only, no commas or units) | b0eaad5a97547afa18d1 | 1,724 | {"log":[{"cursor":null,"items":[{"id":"inv-8716","status":"open","total":278},{"id":"inv-2960","status":"open","total":419},{"id":"inv-4813","status":"paid","total":988}],"next":"cur_5e890","query":"status=any&customer=K-17"},{"cursor":"cur_b770b","items":[{"id":"inv-6293","status":"paid","total":67},{"id":"inv-6709","... |
api_pagination-000013 | api_pagination | agentic_tool_use | applied | 3 | plain | distractor | null | train | api_pagination/plain/p0/s0 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 6 | int | null | exact integer match | an integer (digits only, no commas or units) | 2f32da6923c4e5c5cad5 | 2,472 | {"log":[{"cursor":"cur_c183a","items":[{"id":"inv-7278","status":"paid","total":456}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_78153","items":[{"id":"inv-5377","status":"open","total":252},{"id":"inv-5334","status":"open","total":964},{"id":"inv-2901","status":"open","total":842},{"id":"inv-1475",... |
api_pagination-000014 | api_pagination | agentic_tool_use | applied | 3 | plain | base | null | train | api_pagination/plain/p0/s1 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 1 | int | null | exact integer match | an integer (digits only, no commas or units) | e5f1275aa2977f43bdd8 | 2,296 | {"log":[{"cursor":"cur_39b37","items":[{"id":"inv-2270","status":"paid","total":663}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_7b46e","items":[{"id":"inv-7323","status":"void","total":144},{"id":"inv-2737","status":"open","total":276},{"id":"inv-4332","status":"void","total":288},{"id":"inv-7270",... |
api_pagination-000015 | api_pagination | agentic_tool_use | applied | 3 | plain | base | null | train | api_pagination/plain/p0/s2 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 10 | int | null | exact integer match | an integer (digits only, no commas or units) | 7b9ebc6f9c57a602fc8a | 2,555 | {"log":[{"cursor":"cur_60bda","items":[{"id":"inv-1644","status":"open","total":673},{"id":"inv-3061","status":"paid","total":610},{"id":"inv-2959","status":"open","total":724}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_54a1b","items":[{"id":"inv-9670","status":"paid","total":349}],"next":null,"que... |
api_pagination-000016 | api_pagination | agentic_tool_use | applied | 3 | plain | base | null | train | api_pagination/plain/p0/s3 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 2 | int | null | exact integer match | an integer (digits only, no commas or units) | f85ee2ebaf9d18e33e66 | 2,133 | {"log":[{"cursor":"cur_aab12","items":[{"id":"inv-5349","status":"paid","total":155},{"id":"inv-8811","status":"void","total":11}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_6809e","items":[{"id":"inv-5347","status":"paid","total":49}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_9... |
api_pagination-000019 | api_pagination | agentic_tool_use | applied | 4 | plain | base | null | train | api_pagination/plain/p0/s0 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 5 | int | null | exact integer match | an integer (digits only, no commas or units) | c51482f8a3bce7893344 | 2,749 | {"log":[{"cursor":"cur_d5681","items":[{"id":"inv-2660","status":"paid","total":556},{"id":"inv-1443","status":"void","total":163},{"id":"inv-1618","status":"open","total":604}],"next":"cur_9afdf","query":"status=any&customer=K-17"},{"cursor":"cur_eef22","items":[{"id":"inv-3257","status":"void","total":606},{"id":"inv... |
api_pagination-000020 | api_pagination | agentic_tool_use | applied | 4 | plain | base | null | train | api_pagination/plain/p0/s1 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 8 | int | null | exact integer match | an integer (digits only, no commas or units) | 8555f08b414678619053 | 2,850 | {"log":[{"cursor":"cur_ac720","items":[{"id":"inv-9822","status":"paid","total":368},{"id":"inv-9842","status":"void","total":281},{"id":"inv-3546","status":"void","total":431},{"id":"inv-8778","status":"paid","total":882}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_66068","items":[{"id":"inv-1505",... |
api_pagination-000021 | api_pagination | agentic_tool_use | applied | 4 | plain | base | null | train | api_pagination/plain/p0/s2 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 7 | int | null | exact integer match | an integer (digits only, no commas or units) | 102080517aa40170a0ce | 2,852 | {"log":[{"cursor":null,"items":[{"id":"inv-5877","status":"paid","total":284},{"id":"inv-9858","status":"paid","total":89},{"id":"inv-4328","status":"open","total":955},{"id":"inv-6849","status":"void","total":683}],"next":"cur_3fd5a","query":"status=any&customer=K-17"},{"cursor":"cur_68b2c","items":[{"id":"inv-5815","... |
api_pagination-000022 | api_pagination | agentic_tool_use | applied | 4 | plain | counterfactual_base | pair-api_pagination-000022 | train | api_pagination/plain/p0/s3 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 8 | int | null | exact integer match | an integer (digits only, no commas or units) | 5e8663fee7a1b991e66e | 3,009 | {"log":[{"cursor":"cur_93e3f","items":[{"id":"inv-7623","status":"paid","total":984},{"id":"inv-3580","status":"void","total":69},{"id":"inv-2181","status":"void","total":670},{"id":"inv-3070","status":"open","total":858}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_da504","items":[{"id":"inv-4804","... |
api_pagination-000022-cf | api_pagination | agentic_tool_use | applied | 4 | plain | counterfactual | pair-api_pagination-000022 | train | api_pagination/plain/p0/s3 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 5 | int | null | exact integer match | an integer (digits only, no commas or units) | 747e49d7360bfa1fcef6 | 3,009 | {"log":[{"cursor":"cur_93e3f","items":[{"id":"inv-7623","status":"paid","total":984},{"id":"inv-3580","status":"void","total":69},{"id":"inv-2181","status":"void","total":670},{"id":"inv-3070","status":"open","total":858}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_da504","items":[{"id":"inv-4804","... |
api_pagination-000025 | api_pagination | agentic_tool_use | applied | 5 | plain | base | null | train | api_pagination/plain/p0/s0 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 8 | int | null | exact integer match | an integer (digits only, no commas or units) | 7dc19b7638dd52102b1b | 2,898 | {"log":[{"cursor":null,"items":[{"id":"inv-3294","status":"void","total":816},{"id":"inv-1314","status":"void","total":499}],"next":"cur_dd6e5","query":"status=any&customer=K-17"},{"cursor":"cur_a8fb7","items":[{"id":"inv-6153","status":"paid","total":357}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur... |
api_pagination-000026 | api_pagination | agentic_tool_use | applied | 5 | plain | base | null | train | api_pagination/plain/p0/s1 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 6 | int | null | exact integer match | an integer (digits only, no commas or units) | 49c75b32d8e9259e50c1 | 2,852 | {"log":[{"cursor":"cur_463aa","items":[{"id":"inv-8773","status":"open","total":925},{"id":"inv-8517","status":"open","total":84}],"next":"cur_50fab","query":"status=any&customer=K-17"},{"cursor":null,"items":[{"id":"inv-1326","status":"paid","total":162},{"id":"inv-8443","status":"paid","total":911},{"id":"inv-4216","... |
api_pagination-000027 | api_pagination | agentic_tool_use | applied | 5 | plain | base | null | train | api_pagination/plain/p0/s2 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 4 | int | null | exact integer match | an integer (digits only, no commas or units) | be06e9c7ab7c61c3f913 | 3,008 | {"log":[{"cursor":"cur_a0655","items":[{"id":"inv-9048","status":"paid","total":232},{"id":"inv-1111","status":"void","total":113}],"next":"cur_9595b","query":"status=any&customer=K-17"},{"cursor":"cur_d7e1e","items":[{"id":"inv-8437","status":"paid","total":740},{"id":"inv-8866","status":"void","total":454}],"next":"c... |
api_pagination-000028 | api_pagination | agentic_tool_use | applied | 5 | plain | base | null | train | api_pagination/plain/p0/s3 | Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a different customer's query; a page may appear twice because of a retry). A complete result is obtained by starting from the request without a cursor and following next_cursor until it is null. Invoice IDs a... | 7 | int | null | exact integer match | an integer (digits only, no commas or units) | b01ba1167c7818790598 | 3,045 | {"log":[{"cursor":"cur_7889a","items":[{"id":"inv-1722","status":"paid","total":952}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_ac0ff","items":[{"id":"inv-4956","status":"paid","total":194},{"id":"inv-7014","status":"paid","total":502},{"id":"inv-4334","status":"open","total":975}],"next":"cur_9bae... |
api_pagination-000031 | api_pagination | agentic_tool_use | applied | 1 | scenario | base | null | train | api_pagination/scenario/p0/s0 | You are a integration engineer at Ironbark Health. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a differe... | 5 | int | null | exact integer match | an integer (digits only, no commas or units) | aa8d62303350649dab8a | 1,574 | {"log":[{"cursor":"cur_a6b8d","items":[{"id":"inv-7777","status":"void","total":364},{"id":"inv-3957","status":"paid","total":933},{"id":"inv-1082","status":"open","total":669},{"id":"inv-3432","status":"paid","total":338}],"next":"cur_6a622","query":"status=any&customer=K-17"},{"cursor":"cur_e8acb","items":[{"id":"inv... |
api_pagination-000032 | api_pagination | agentic_tool_use | applied | 1 | scenario | base | null | train | api_pagination/scenario/p0/s1 | You are a senior integration engineer at Helios Grid Services. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong ... | 2 | int | null | exact integer match | an integer (digits only, no commas or units) | 1d5d6dca771ad09e5c5e | 1,586 | {"log":[{"cursor":"cur_9ff1f","items":[{"id":"inv-7116","status":"void","total":666},{"id":"inv-1524","status":"open","total":354},{"id":"inv-2151","status":"paid","total":119},{"id":"inv-9276","status":"open","total":977}],"next":"cur_b325c","query":"status=any&customer=K-17"},{"cursor":"cur_7873e","items":[{"id":"inv... |
api_pagination-000033 | api_pagination | agentic_tool_use | applied | 1 | scenario | counterfactual_base | pair-api_pagination-000033 | train | api_pagination/scenario/p0/s2 | You are a integration engineer on call at Nimbus Cloudworks. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to... | 3 | int | null | exact integer match | an integer (digits only, no commas or units) | 079711eae9b3966cfe6f | 1,584 | {"log":[{"cursor":"cur_a0fc3","items":[{"id":"inv-9787","status":"void","total":703},{"id":"inv-6206","status":"void","total":452},{"id":"inv-4087","status":"void","total":963},{"id":"inv-4738","status":"paid","total":674}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_4771e","items":[{"id":"inv-7630",... |
api_pagination-000033-cf | api_pagination | agentic_tool_use | applied | 1 | scenario | counterfactual | pair-api_pagination-000033 | train | api_pagination/scenario/p0/s2 | You are a integration engineer on call at Nimbus Cloudworks. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to... | 1 | int | null | exact integer match | an integer (digits only, no commas or units) | df731166def6a96b5381 | 1,584 | {"log":[{"cursor":"cur_a0fc3","items":[{"id":"inv-9787","status":"void","total":703},{"id":"inv-6206","status":"void","total":452},{"id":"inv-4087","status":"void","total":963},{"id":"inv-4738","status":"paid","total":674}],"next":null,"query":"status=any&customer=K-17"},{"cursor":"cur_4771e","items":[{"id":"inv-7630",... |
api_pagination-000034 | api_pagination | agentic_tool_use | applied | 1 | scenario | counterfactual_base | pair-api_pagination-000034 | train | api_pagination/scenario/p0/s3 | You are a lead integration engineer at Kestrel Freight. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a di... | 4 | int | null | exact integer match | an integer (digits only, no commas or units) | cb2d18d5c3eaefdb8e79 | 1,735 | {"log":[{"cursor":"cur_24e87","items":[{"id":"inv-7910","status":"void","total":335},{"id":"inv-5740","status":"open","total":411},{"id":"inv-5628","status":"paid","total":961}],"next":"cur_a0134","query":"status=any&customer=K-17"},{"cursor":null,"items":[{"id":"inv-5905","status":"paid","total":398},{"id":"inv-3111",... |
api_pagination-000034-cf | api_pagination | agentic_tool_use | applied | 1 | scenario | counterfactual | pair-api_pagination-000034 | train | api_pagination/scenario/p0/s3 | You are a lead integration engineer at Kestrel Freight. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a di... | 5 | int | null | exact integer match | an integer (digits only, no commas or units) | a967c21800b42c3b46e7 | 1,735 | {"log":[{"cursor":"cur_24e87","items":[{"id":"inv-7910","status":"void","total":335},{"id":"inv-5740","status":"open","total":411},{"id":"inv-5628","status":"paid","total":961}],"next":"cur_a0134","query":"status=any&customer=K-17"},{"cursor":null,"items":[{"id":"inv-5905","status":"paid","total":398},{"id":"inv-3111",... |
api_pagination-000037 | api_pagination | agentic_tool_use | applied | 2 | scenario | counterfactual_base | pair-api_pagination-000037 | train | api_pagination/scenario/p0/s0 | You are a integration engineer at Ironbark Health. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a differe... | 2 | int | null | exact integer match | an integer (digits only, no commas or units) | 7771372e0110c5d3462c | 2,081 | {"log":[{"cursor":"cur_e0d1e","items":[{"id":"inv-5326","status":"paid","total":189},{"id":"inv-3476","status":"void","total":772},{"id":"inv-4445","status":"open","total":40},{"id":"inv-8152","status":"void","total":134}],"next":"cur_d3c29","query":"status=any&customer=K-17"},{"cursor":"cur_6a32f","items":[{"id":"inv-... |
api_pagination-000037-cf | api_pagination | agentic_tool_use | applied | 2 | scenario | counterfactual | pair-api_pagination-000037 | train | api_pagination/scenario/p0/s0 | You are a integration engineer at Ironbark Health. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a differe... | 5 | int | null | exact integer match | an integer (digits only, no commas or units) | 5681025840dc5ba951e9 | 2,081 | {"log":[{"cursor":"cur_e0d1e","items":[{"id":"inv-5326","status":"paid","total":189},{"id":"inv-3476","status":"void","total":772},{"id":"inv-4445","status":"open","total":40},{"id":"inv-8152","status":"void","total":134}],"next":"cur_d3c29","query":"status=any&customer=K-17"},{"cursor":"cur_6a32f","items":[{"id":"inv-... |
api_pagination-000038 | api_pagination | agentic_tool_use | applied | 2 | scenario | base | null | train | api_pagination/scenario/p0/s1 | You are a senior integration engineer at Helios Grid Services. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong ... | 2 | int | null | exact integer match | an integer (digits only, no commas or units) | a70b467573bd897f6bd2 | 2,147 | {"log":[{"cursor":null,"items":[{"id":"inv-4535","status":"open","total":199},{"id":"inv-2400","status":"void","total":848},{"id":"inv-7802","status":"paid","total":456},{"id":"inv-7260","status":"void","total":401}],"next":"cur_764d3","query":"status=any&customer=K-17"},{"cursor":"cur_764d3","items":[{"id":"inv-8953",... |
api_pagination-000039 | api_pagination | agentic_tool_use | applied | 2 | scenario | distractor | null | train | api_pagination/scenario/p0/s2 | You are a integration engineer on call at Nimbus Cloudworks. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to... | 6 | int | null | exact integer match | an integer (digits only, no commas or units) | f22a552ffd61945bd10e | 2,173 | {"log":[{"cursor":"cur_63621","items":[{"id":"inv-2021","status":"paid","total":843}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_44eea","items":[{"id":"inv-8676","status":"void","total":939},{"id":"inv-8163","status":"void","total":193},{"id":"inv-2287","status":"paid","total":785}],"next":null,"que... |
api_pagination-000040 | api_pagination | agentic_tool_use | applied | 2 | scenario | base | null | train | api_pagination/scenario/p0/s3 | You are a lead integration engineer at Kestrel Freight. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a di... | 4 | int | null | exact integer match | an integer (digits only, no commas or units) | 4aa027c1c1995b9b4710 | 1,982 | {"log":[{"cursor":"cur_58d6f","items":[{"id":"inv-9222","status":"open","total":45},{"id":"inv-7800","status":"void","total":402},{"id":"inv-3360","status":"paid","total":254},{"id":"inv-3299","status":"void","total":467}],"next":"cur_43f89","query":"status=any&customer=K-17"},{"cursor":"cur_1f800","items":[{"id":"inv-... |
api_pagination-000043 | api_pagination | agentic_tool_use | applied | 3 | scenario | counterfactual_base | pair-api_pagination-000043 | train | api_pagination/scenario/p0/s0 | You are a integration engineer at Ironbark Health. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a differe... | 2 | int | null | exact integer match | an integer (digits only, no commas or units) | f9317a0c11ad32192ea0 | 2,494 | {"log":[{"cursor":"cur_edfa9","items":[{"id":"inv-8954","status":"paid","total":795}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_dd4cf","items":[{"id":"inv-9978","status":"void","total":387},{"id":"inv-4330","status":"paid","total":418}],"next":"cur_9de16","query":"status=any&customer=K-17"},{"curso... |
api_pagination-000043-cf | api_pagination | agentic_tool_use | applied | 3 | scenario | counterfactual | pair-api_pagination-000043 | train | api_pagination/scenario/p0/s0 | You are a integration engineer at Ironbark Health. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a differe... | 6 | int | null | exact integer match | an integer (digits only, no commas or units) | a18012b246feacd07cb2 | 2,494 | {"log":[{"cursor":"cur_edfa9","items":[{"id":"inv-8954","status":"paid","total":795}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_dd4cf","items":[{"id":"inv-9978","status":"void","total":387},{"id":"inv-4330","status":"paid","total":418}],"next":"cur_9de16","query":"status=any&customer=K-17"},{"curso... |
api_pagination-000044 | api_pagination | agentic_tool_use | applied | 3 | scenario | base | null | train | api_pagination/scenario/p0/s1 | You are a senior integration engineer at Helios Grid Services. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong ... | 5 | int | null | exact integer match | an integer (digits only, no commas or units) | c6a462b4245d36900dc7 | 2,455 | {"log":[{"cursor":"cur_90b37","items":[{"id":"inv-4924","status":"open","total":850},{"id":"inv-7763","status":"paid","total":833}],"next":"cur_c3444","query":"status=any&customer=K-17"},{"cursor":"cur_90b37","items":[{"id":"inv-4924","status":"open","total":850},{"id":"inv-7763","status":"paid","total":833}],"next":"c... |
api_pagination-000045 | api_pagination | agentic_tool_use | applied | 3 | scenario | counterfactual_base | pair-api_pagination-000045 | train | api_pagination/scenario/p0/s2 | You are a integration engineer on call at Nimbus Cloudworks. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to... | 3 | int | null | exact integer match | an integer (digits only, no commas or units) | 44fa68cfae64d6d4e65c | 2,487 | {"log":[{"cursor":"cur_43f0f","items":[{"id":"inv-3600","status":"void","total":570},{"id":"inv-9106","status":"paid","total":337}],"next":"cur_e9e4c","query":"status=any&customer=K-17"},{"cursor":"cur_221c6","items":[{"id":"inv-3355","status":"open","total":957},{"id":"inv-3356","status":"open","total":524}],"next":"c... |
api_pagination-000045-cf | api_pagination | agentic_tool_use | applied | 3 | scenario | counterfactual | pair-api_pagination-000045 | train | api_pagination/scenario/p0/s2 | You are a integration engineer on call at Nimbus Cloudworks. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to... | 7 | int | null | exact integer match | an integer (digits only, no commas or units) | c1b71fae6fef1abae662 | 2,487 | {"log":[{"cursor":"cur_43f0f","items":[{"id":"inv-3600","status":"void","total":570},{"id":"inv-9106","status":"paid","total":337}],"next":"cur_e9e4c","query":"status=any&customer=K-17"},{"cursor":"cur_221c6","items":[{"id":"inv-3355","status":"open","total":957},{"id":"inv-3356","status":"open","total":524}],"next":"c... |
api_pagination-000046 | api_pagination | agentic_tool_use | applied | 3 | scenario | base | null | train | api_pagination/scenario/p0/s3 | You are a lead integration engineer at Kestrel Freight. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a di... | 4 | int | null | exact integer match | an integer (digits only, no commas or units) | a191f1b4a3648c3a016c | 2,482 | {"log":[{"cursor":"cur_42e3c","items":[{"id":"inv-7695","status":"paid","total":912}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_b07e9","items":[{"id":"inv-5978","status":"void","total":137},{"id":"inv-8522","status":"void","total":545}],"next":"cur_ca0a8","query":"status=any&customer=K-17"},{"curso... |
api_pagination-000049 | api_pagination | agentic_tool_use | applied | 4 | scenario | distractor | null | train | api_pagination/scenario/p0/s0 | You are a integration engineer at Ironbark Health. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong to a differe... | 4 | int | null | exact integer match | an integer (digits only, no commas or units) | 3b1cca28590060e36a27 | 3,022 | {"log":[{"cursor":null,"items":[{"id":"inv-1524","status":"void","total":865},{"id":"inv-7238","status":"paid","total":394}],"next":"cur_9d652","query":"status=any&customer=K-17"},{"cursor":"cur_1f8a9","items":[{"id":"inv-3858","status":"paid","total":876}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur... |
api_pagination-000050 | api_pagination | agentic_tool_use | applied | 4 | scenario | base | null | train | api_pagination/scenario/p0/s1 | You are a senior integration engineer at Helios Grid Services. A colleague has asked you to reconstruct a paginated API result from a messy request log. Work carefully; the result goes into an internal report.
Request log of an integration fetching all invoices for customer K-17 (entries are out of order; some belong ... | 5 | int | null | exact integer match | an integer (digits only, no commas or units) | d367547b074af5b8a2ef | 2,911 | {"log":[{"cursor":"cur_25011","items":[{"id":"inv-7084","status":"paid","total":853}],"next":null,"query":"status=any&customer=K-42"},{"cursor":"cur_67d4f","items":[{"id":"inv-3588","status":"void","total":615},{"id":"inv-2241","status":"paid","total":548},{"id":"inv-9528","status":"void","total":574}],"next":"cur_a33a... |
End of preview. Expand in Data Studio
Beam Verified Reasoning v7
by Marshoot Lab
63,620 freshly generated reasoning problems across 58 task families and 17 domains, each with an answer computed by code and recomputed by a second, independent implementation, plus 1,241 graded reasoning traces from Reflection's Beam-1.0-501B-A23B at low and high reasoning effort, and a held-out evaluation of Beam against 3 other openly available models on the same items.
90.4% of the problems are applied, professional-style tasks (policy compliance, ledger audits, defensive log forensics, configuration reasoning, tool-use transcripts, long multi-hop documents, simulations, data wrangling); classic math and logic puzzles are kept to 9.6% and appear only at hard difficulty tiers.
Highlights
| Verified problems | 63,620 (42,627 train / 10,528 validation / 10,465 test) |
| Task families / domains | 58 / 17, 5 difficulty levels and 4 prompt styles per family |
| Key verification | two independent implementations per family; 0 disagreements in the released rows (74,542 candidate instances generated) |
| Leakage | split by prompt-template group; 0 problems, 0 template groups and 0 seeds shared across splits |
| Near-duplicates | MinHash check over all rows; 1,185 rows removed, including every cross-split near-duplicate |
| Adversarial variants | 5,057 distractor items and 2,393 counterfactual pairs (one constraint changed, different key) |
| Beam traces | 1,241 traces (high: 812, low: 429), 1,170 verified correct, 71 incorrect kept and labelled |
| Held-out evaluation | 212 shared test items; Beam high 95.8%, Beam low 90.6% |
Task taxonomy
| Domain | Families | Problems | Train / Val / Test | What the tasks look like |
|---|---|---|---|---|
| Rule/policy compliance | 6 | 6,900 | 4,606 / 1,146 / 1,148 | expense policy with caps and receipt rules, carrier rate card, RBAC/ABAC access policy with deny-overrides, fictional tax code with phase-outs, leave accrual, SLA credits |
| Spreadsheet / ledger auditing | 3 | 3,451 | 2,303 / 575 / 573 | first mis-keyed running balance, bank reconciliation, invoice line audit |
| Log & incident forensics (defensive) | 3 | 3,450 | 2,298 / 573 / 579 | brute-force detection windows, multi-source intrusion timeline (DNS + firewall + logon, mixed UTC offsets), DNS beaconing |
| Cyber defense | 5 | 5,708 | 3,814 / 943 / 951 | first-match firewall evaluation, subnet plans, POSIX permissions and chmod, CVSS v3.1 base score, encoding chains |
| Code reasoning | 4 | 4,591 | 3,069 / 765 / 757 | program output tracing, bug localisation from a failing test, SQL results on inline tables, regex acceptance |
| State machines & simulation | 6 | 6,695 | 4,495 / 1,113 / 1,087 | FIFO inventory with backorders and returns, bank account with fees and interest, service queue, elevator (LOOK), robot grid, vending FSM |
| Long-context multi-hop | 3 | 3,450 | 2,305 / 574 / 571 | directory memo with look-alike names and superseded facts, contract with amendments and a void side letter, action items across meeting notes |
| Tool-use transcripts | 2 | 2,301 | 1,537 / 381 / 383 | audit an agent transcript with wrong-argument calls, reconstruct a paginated API result from an out-of-order log |
| Data wrangling | 4 | 4,600 | 3,078 / 761 / 761 | multi-currency export with mixed date/number formats and duplicates, record de-duplication, mixed-unit sensor feed, nested JSON aggregation |
| Configuration reasoning | 6 | 6,815 | 4,556 / 1,137 / 1,122 | cron next-fire times, semver ranges, dependency resolution, Kubernetes requests, time zones across DST, business-day SLAs |
| Scheduling & routing | 3 | 3,451 | 2,303 / 580 / 568 | critical path / float, cheapest route with closures and tolls, meeting rooms |
| Finance | 2 | 2,279 | 1,533 / 374 / 372 | loan amortisation with extra payments, NPV and payback |
| Statistics | 2 | 2,171 | 1,482 / 346 / 343 | percentiles, sample SD, IQR outliers, base-rate (Bayes) reasoning |
| Lab chemistry | 1 | 671 | 497 / 84 / 90 | reagent mass with purity and hydrates, dilution |
| Units | 1 | 951 | 648 / 155 / 148 | compound unit conversion with exact definitions |
| Classic math (hard tiers only) | 5 | 4,136 | 2,767 / 689 / 680 | CRT / modular exponentiation, constrained arrangements, exact dice probabilities, linear systems, lattice polygons |
| Classic logic (hard tiers only) | 2 | 2,000 | 1,336 / 332 / 332 | boolean circuits, house/pet constraint puzzles with unique solutions |
| Total | 58 | 63,620 | 42,627 / 10,528 / 10,465 |
Every family has five difficulty levels (applied families use levels 1-5; classic families use levels 4-8 of the same scale, so they contribute hard items only) and four prompt styles: plain, scenario (a short workplace framing), structured (all inputs as JSON) and chained (two sub-questions over the same input, combined into one final answer). Each prompt states the exact answer format, and each row carries a grading_rule.
Why these task types
The other dataset entries in the Build with Beam thread concentrate on a few formats. We measured their published files (review dated 2026-10-27):
- The largest, ProCreations/Beam-Reasoning-Traces (53,306 SFT rows), contains two sources only: DeepMath-103K math (29,578 rows) and OpenScienceReasoning-2 science (23,728 rows), and 24,020 of its rows (45%) have a single-letter multiple-choice key. Its keys come from those public datasets.
- BeamDial's 410 tasks are all from public benchmarks (GSM8K, MATH, AMC23, AIME24, OlympiadBench, HumanEval, MBPP), with Beam at 96-97% at every effort; the author flags 6 of the reference keys as wrong or underspecified.
- RulebookBench (11,192 train rows) is carefully built (we recomputed all 9,410 of its derivation labels and found 0 errors), but every item is a Horn-clause derivation with a binary label, 50% of the training rows use an abstract
datalogskin, and none asks for an amount, date or count. - The Effort Atlas has 36 tasks and 206 traces, of which 45 ended at the token limit; RegexLogic-RLVR has 16 prompts.
No entry provides ledger reconciliation, defensive log timelines, configuration reasoning (cron, semver, DST, Kubernetes), tool-transcript auditing or long multi-hop documents as plain prompt/answer rows with computed keys, so v7 centres on those.
How the answers are verified
For every family we wrote the answer twice, with different methods, and keep an instance only when both agree:
| Family examples | Implementation A | Implementation B |
|---|---|---|
| SQL on inline tables | executes the query in SQLite | re-implements the query with pandas |
| Time zones / DST | zoneinfo (IANA database) |
hand-written US / EU / AU DST rules |
| Regex acceptance | Python re.fullmatch |
our own backtracking matcher over the pattern's syntax tree |
| Firewall / subnets | ipaddress module |
integer bit-mask arithmetic |
| Business-day SLAs | day-by-day calendar loop | numpy.busday_offset with a holiday calendar |
| Critical path | forward / backward pass | longest path on an activity graph (networkx) |
| Code tracing | executes the generated program in a restricted interpreter | interprets the program's specification without running the source |
| Bug localisation | the line that was mutated | searches every single-line edit and checks which one makes a test battery pass (must be exactly one line) |
| Money, tax, interest | exact Fraction arithmetic |
Decimal arithmetic with half-up rounding |
| Dedup / clustering | union-find | graph connected components (networkx) |
During development the second implementation caught real bugs in the first: wall-clock arithmetic across a DST change (a flight's arrival was off by one hour), a regex matcher that mishandled empty repetitions, polygons that became self-intersecting after rounding, and non-unique invoice IDs in a pagination task. Each was fixed and the families regenerated; the released rows contain 0 disagreements between implementations. Instances whose key would be ambiguous (ties, rounding-boundary values, puzzles without a unique solution) are discarded (25 during the final build). Each row's metadata field holds the generated world state, so keys can be re-derived independently.
Quality gates
| Gate | Result |
|---|---|
| Implementation A vs B disagreements in released rows | 0 |
| Grader self-check (every key graded against itself) | 0 failures |
| Exact duplicate prompts (normalised) | 8 removed |
| Duplicate underlying problems | skipped at generation (12,156 candidates) |
| Near-duplicates (MinHash, est. Jaccard >= 0.9) | 1,363 within-split and 859 cross-split pairs; 1,185 rows removed (held-out copy dropped for every cross-split pair) |
| Problems / template groups / seeds shared across splits | 0 / 0 / 0 |
| Counterfactual pairs split across splits | 0 |
| Non-Latin script characters | 0 rows |
| Prompts without an explicit final-answer format | 0 |
| Prompt length (characters) | min 184, median 1,061, max 13,815 |
| PII | synthetic names only; e-mail addresses on reserved example domains; phone numbers in the fictional 555-01xx range; external IPs only from the RFC 5737 documentation ranges. The scan found 0 formatted phone numbers outside 555-01xx; its 10-digit hits are moduli in number-theory items and test strings in regex items. |
Splits. Each family has six workplace "skins" (organisation, role and framing) crossed with two phrasings. All items of skin 5 form the test split and all items of skin 4 the validation split, so no prompt template group appears in more than one split. Generator seeds are also disjoint. For the plain, structured and chained styles the skins differ less in wording than in the scenario style, so for those the held-out unit is mainly the generated instance.
Beam reasoning traces
Traces were generated by querying Beam-1.0-501B-A23B through a hosted API with reasoning_effort low and high, temperature 0.7, top_p 0.99 and a 16,384-token output limit, with the prompt as the only user message. All Beam numbers below come from one serving path. Every trace is graded by the same rule-based grader against the computed key; incorrect traces are kept and labelled (is_correct, verified_correct), which makes them usable for preference and RL data.
| Low effort | High effort | |
|---|---|---|
| Traces | 429 | 812 |
| Accuracy (all traced items) | 90.7% | 96.2% |
| Median completion tokens | 409 | 1,112 |
| Domain | Low effort (n) | High effort (n) |
|---|---|---|
| finance | 37.5% (16) | 96.0% (25) |
| chemistry | 62.5% (8) | 52.9% (17) |
| classic math | 75.9% (29) | 87.3% (63) |
| cyber defense | 81.6% (38) | 94.0% (67) |
| code reasoning | 89.3% (28) | 96.4% (56) |
| classic logic | 91.7% (12) | 100.0% (28) |
| policy compliance | 93.9% (49) | 97.8% (90) |
| ledger audit | 95.5% (22) | 100.0% (40) |
| config reasoning | 95.8% (48) | 100.0% (90) |
| scheduling | 95.8% (24) | 100.0% (44) |
| state simulation | 95.8% (48) | 95.5% (89) |
| physics units | 100.0% (6) | 100.0% (13) |
| agentic tool use | 100.0% (16) | 100.0% (31) |
| log forensics | 100.0% (20) | 100.0% (45) |
| data wrangling | 100.0% (31) | 100.0% (50) |
| long context | 100.0% (20) | 97.4% (39) |
| statistics | 100.0% (14) | 96.0% (25) |
By difficulty level:
| Difficulty | Low effort (n) | High effort (n) |
|---|---|---|
| 1 | 93.3% (75) | 97.2% (145) |
| 2 | 92.3% (78) | 95.6% (135) |
| 3 | 90.0% (70) | 97.8% (139) |
| 4 | 89.8% (88) | 98.8% (165) |
| 5 | 92.4% (92) | 94.8% (172) |
| 6 | 77.8% (9) | 83.3% (18) |
| 7 | 85.7% (7) | 93.3% (15) |
| 8 | 70.0% (10) | 87.0% (23) |
When does high effort help? On the 425 held-out evaluation items run at both efforts, low effort scored 90.8% and high effort 96.7%. High effort fixed 31 low-effort errors and broke 6 low-effort successes (exact McNemar p = 4.13e-05), at 3.5x the mean output tokens (528 vs 1,861). A simple policy that re-asks at high effort only when the low-effort answer used more than the median number of tokens (410) reached 96.0% at 2,025 mean tokens per item (low and high tokens both counted for escalated items).
Evaluation
Setup. 212 items from the test split, shared by every model in the table: one base item per family and difficulty level, plus one distractor item and one counterfactual pair per family where available, taken in a stratified round-robin order. Every model received the identical prompt (no system prompt), temperature 0.7, top_p 0.99 and a 16,384-token output limit; Beam was run at low and high effort and the other models at their default reasoning setting. Grading is programmatic only (the final Final answer: line, checked against the key with the row's grading rule). All models were queried through hosted free-tier services in October 2026; these deployments may serve quantized weights and are rate-limited, so results describe these deployments rather than the models' best possible configurations. Intervals are 95% Wilson intervals; paired comparisons use the exact McNemar test.
| Model | Accuracy (95% CI) | Format compliance | Mean output tokens | Accuracy per 1k tokens | Median latency (s) |
|---|---|---|---|---|---|
| Beam-1.0-501B-A23B (high) | 95.8% (92.1-97.8) | 95.3% | 2,107 | 0.45 | 9.3 |
| gpt-oss-120b | 92.5% (88.1-95.3) | 95.8% | 1,531 | 0.60 | 6.8 |
| Beam-1.0-501B-A23B (low) | 90.6% (85.9-93.8) | 97.2% | 557 | 1.63 | 4.5 |
| Nemotron-3-Super | 89.6% (84.8-93.0) | 92.5% | 3,737 | 0.24 | 20.6 |
| Gemma-4-31B | 88.2% (83.2-91.9) | 96.7% | 798 | 1.11 | 5.5 |
Models with fewer completed items are listed below; their numbers cover only the items they finished (the first items of the same stratified order) and are not comparable to the table above.
| Model | Items completed | Accuracy (95% CI) | Truncated |
|---|---|---|---|
| Beam-1.0-501B-A23B (high) | 425 | 96.7% (94.5-98.0) | 4 |
| gpt-oss-120b | 425 | 92.0% (89.0-94.2) | 2 |
| Beam-1.0-501B-A23B (low) | 425 | 90.8% (87.7-93.2) | 0 |
| Gemma-4-31B | 425 | 89.6% (86.4-92.2) | 0 |
| Nemotron-3-Super | 212 | 89.6% (84.8-93.0) | 10 |
| Laguna-S-2.1 | 12 | 75.0% (46.8-91.1) | 0 |
| dots-3-note-preview | 11 | 90.9% (62.3-98.4) | 0 |
| Nemotron-3-Ultra-550B-A55B | 8 | 100.0% (67.6-100.0) | 0 |
| Flint-v1.0 | 8 | 75.0% (40.9-92.9) | 1 |
| Nemotron-3-Nano-Omni-30B-A3B | 8 | 75.0% (40.9-92.9) | 0 |
| LFM-2.5-2.6B | 4 | 75.0% (30.1-95.4) | 0 |
| Nemotron-3.5-Lightning | 3 | 100.0% (43.8-100.0) | 0 |
| Laguna-XS-2.1 | 2 | 100.0% (34.2-100.0) | 0 |
| Xing-4.0-29B-A4B | 1 | 100.0% (20.7-100.0) | 0 |
Adversarial and counterfactual items. Distractor items add irrelevant but plausible notes; a counterfactual pair is the same problem with one constraint changed so that the key changes. "Answered with the unchanged problem's key" counts counterfactual items where the model gave the base problem's answer.
| Model | Base items (n, acc) | Distractor items (n, acc) | Counterfactual pairs both correct (n, rate) | Answered with the unchanged problem's key |
|---|---|---|---|---|
| Beam-1.0-501B-A23B (high) | 290, 96.6% | 57, 98.2% | 39, 94.9% | 0 |
| Nemotron-3-Super | 128, 93.0% | 57, 87.7% | 13, 69.2% | 1 |
| gpt-oss-120b | 290, 91.4% | 57, 94.7% | 39, 87.2% | 0 |
| Beam-1.0-501B-A23B (low) | 290, 91.0% | 57, 91.2% | 39, 79.5% | 0 |
| Gemma-4-31B | 290, 89.3% | 57, 89.5% | 39, 87.2% | 0 |
Failure taxonomy (incorrect answers on all completed items, classified programmatically in this order: truncated or empty, no extractable final answer, counterfactual answered with the base key, numeric answer within 1% of the key, wrong on a distractor item, other wrong reasoning):
| Model | wrong reasoning / rule misread | near-miss numeric (<=1% off) | wrong on distractor item | ignored counterfactual change | format error (no final answer) | truncated / empty |
|---|---|---|---|---|---|---|
| Beam-1.0-501B-A23B (high) | 10 | 0 | 0 | 0 | 0 | 4 |
| Beam-1.0-501B-A23B (low) | 27 | 8 | 4 | 0 | 0 | 0 |
| Gemma-4-31B | 25 | 15 | 4 | 0 | 0 | 0 |
| Nemotron-3-Super | 4 | 3 | 4 | 1 | 0 | 10 |
| gpt-oss-120b | 19 | 12 | 1 | 0 | 0 | 2 |
To check the classification and the keys, we read 14 randomly sampled incorrect answers by hand. In every case we could recompute, the key was confirmed and the error was the model's: dropping the -0.029 term in the CVSS scope-changed impact formula, arithmetic drift in long amortisation schedules, times given without the date, a unit conversion rounded to 3 instead of 4 significant figures, a degenerate output of repeated <unk> tokens, and responses cut off at the 16,384-token limit.
Paired comparison with Beam at high effort (same items):
| Beam (high) vs | n | Only Beam correct | Only other correct | Exact McNemar p |
|---|---|---|---|---|
| Gemma-4-31B | 212 | 20 | 4 | 0.00154 |
| Nemotron-3-Super | 212 | 18 | 5 | 0.0106 |
| Beam-1.0-501B-A23B (low) | 212 | 15 | 4 | 0.0192 |
| gpt-oss-120b | 212 | 12 | 5 | 0.143 |
On the shared items, Beam at high effort was the only model to answer 7 items correctly, and the only model to get 3 items wrong.
Fine-tuning proof
To check that the training split teaches skills that transfer to unseen templates, we fine-tuned open instruct models on it and compared each with its untouched base model on the same held-out items.
- Qwen2.5-1.5B-Instruct (Apache-2.0): LoRA SFT (bf16, r=32, all linear layers, 1 epoch) on train-split answer keys; 8,282 training problems, 549 optimizer steps (30.8 min); one NVIDIA A10G (24 GB) GPU, about 46 GPU-minutes in total, roughly $0.77.
- Qwen2.5-7B-Instruct (Apache-2.0): LoRA SFT (bf16, r=16, all linear layers, 1 epoch) on train-split answer keys; 19,714 training problems, 636 optimizer steps (62.9 min); one NVIDIA A100 (80 GB) GPU, about 85 GPU-minutes in total, roughly $3.56.
- Training data: problems sampled from the
trainsplit across all 58 families; target =Final answer: <canonical answer>(answer keys only, no written reasoning and no Beam traces). No test-split rows and no shared template groups or problem hashes were used in training. - Evaluation: the same 425 stratified
test-split items for every model (a fixed set chosen before training, also the pool for the Evaluation section), greedy, max_new_tokens=1024, identical prompt, graded programmatically with the dataset's own answer types and tolerances.
| Model | n | Accuracy | 95% Wilson CI | Format-compliant | Median output tokens | Paired (FT-only / base-only), McNemar p |
|---|---|---|---|---|---|---|
| Qwen2.5-1.5B-Instruct base | 425 | 4.9% (21/425) | 3.2% – 7.4% | 68.7% | 15 | |
| Qwen2.5-1.5B-Instruct + LoRA SFT | 425 | 15.3% (65/425) | 12.2% – 19.0% | 99.8% | 9 | 53 / 9, p = 1.0e-08 |
| Qwen2.5-7B-Instruct base | 425 | 21.2% (90/425) | 17.6% – 25.3% | 77.6% | 525 | |
| Qwen2.5-7B-Instruct + LoRA SFT | 425 | 26.6% (113/425) | 22.6% – 31.0% | 100.0% | 10 | 66 / 43, p = 3.5e-02 |
- Qwen2.5-1.5B-Instruct: on the 292 items where the base model's answer was already in the required format, the base model solved 20 and the fine-tuned model 51.
- Qwen2.5-7B-Instruct: on the 330 items where the base model's answer was already in the required format, the base model solved 82 and the fine-tuned model 96.
The gain shrinks with a stronger base model: +10.4 pts for Qwen2.5-1.5B-Instruct (p = 1.0e-08) versus +5.4 pts for Qwen2.5-7B-Instruct (p = 0.03), and for Qwen2.5-7B-Instruct the confidence intervals of base and fine-tuned overlap.
| Domain | n | 7B base | 7B FT | 7B Δ (pts) | 1.5B base | 1.5B FT | 1.5B Δ (pts) |
|---|---|---|---|---|---|---|---|
| log_forensics | 20 | 10.0% | 50.0% | +40.0 | 5.0% | 15.0% | +10.0 |
| cyber_defense | 36 | 27.8% | 55.6% | +27.8 | 11.1% | 16.7% | +5.6 |
| code_reasoning | 28 | 10.7% | 28.6% | +17.9 | 0.0% | 10.7% | +10.7 |
| long_context | 20 | 35.0% | 50.0% | +15.0 | 20.0% | 40.0% | +20.0 |
| config_reasoning | 48 | 16.7% | 27.1% | +10.4 | 0.0% | 18.8% | +18.8 |
| classic_logic | 12 | 33.3% | 41.7% | +8.3 | 33.3% | 33.3% | +0.0 |
| data_wrangling | 30 | 33.3% | 40.0% | +6.7 | 0.0% | 23.3% | +23.3 |
| finance | 16 | 0.0% | 6.2% | +6.2 | 0.0% | 0.0% | +0.0 |
| policy_compliance | 48 | 16.7% | 20.8% | +4.2 | 6.2% | 22.9% | +16.7 |
| state_simulation | 48 | 12.5% | 16.7% | +4.2 | 4.2% | 12.5% | +8.3 |
| agentic_tool_use | 16 | 31.2% | 31.2% | +0.0 | 0.0% | 18.8% | +18.8 |
| classic_math | 29 | 20.7% | 13.8% | -6.9 | 3.5% | 3.5% | +0.0 |
| ledger_audit | 22 | 27.3% | 13.6% | -13.6 | 4.5% | 0.0% | -4.5 |
| statistics | 14 | 21.4% | 7.1% | -14.3 | 0.0% | 7.1% | +7.1 |
| scheduling | 24 | 33.3% | 12.5% | -20.8 | 0.0% | 12.5% | +12.5 |
| chemistry | 8 | 25.0% | 0.0% | -25.0 | 0.0% | 0.0% | +0.0 |
| physics_units | 6 | 33.3% | 0.0% | -33.3 | 16.7% | 0.0% | -16.7 |
Caveats: each model was trained once with one seed. The fine-tuned models were trained on verified answer keys and answer without written reasoning, while the base models usually reason before answering, so this measures how well the split transfers to unseen templates, not reasoning quality. Per-domain samples are small (6–48 items), so per-domain deltas carry wide intervals; for Qwen2.5-7B-Instruct, 1 domain is flat and 6 dropped (physics units, chemistry, scheduling, statistics, ledger audit, classic math). The drops are concentrated in calculation-heavy domains, where the base model's step-by-step reasoning (median 525 output tokens vs 10 after fine-tuning) likely helps; training on traces that include the reasoning, rather than answer keys alone, is the obvious next step. All models had the same 1024-token budget (Qwen2.5-1.5B-Instruct: 30 base / 1 fine-tuned outputs hit it; Qwen2.5-7B-Instruct: 47 base / 1 fine-tuned outputs hit it); a larger budget would likely help the reasoning base models somewhat more.
Comparison with other Build with Beam entries
| This dataset (v7) | ProCreations/Beam-Reasoning-Traces | RulebookBench | BeamDial | Beam Effort Atlas | |
|---|---|---|---|---|---|
| Problems | 63,620 generated | 53,306 SFT prompts from DeepMath-103K and OpenScienceReasoning-2 | 11,192 train + 1,120 dev + 4,220 test, generated | 410 from public benchmarks | 36 |
| Domains | 17 (58 families) | math, science | rule derivation (5 skins) | math, code | 21 task families |
| Key source | computed twice by independent code | inherited from the source datasets | generator (0 errors in our recomputation) | benchmark references (6 flagged by the author) | computed |
| Multiple-choice keys | none | 24,020 rows | binary labels | none | none |
| Beam traces | 1,241 (low and high) | 124,945 generations, mostly medium effort | 2,160 result rows | 2,050 | 206 |
| Other models on the same items | 3 | own distilled models | Qwen | none | none |
Figures for other entries are from their published files and cards as of 2026-10-27. ProCreations' dataset has many more Beam traces than ours; ours has more problem rows, more domains and dual-implementation keys.
Data fields
problems (configs all and one per domain): id, family, domain, category (applied / classic), difficulty, prompt_style, variant (base / distractor / counterfactual_base / counterfactual), pair_id, split, template_group, prompt, answer, answer_type, answer_part_types, grading_rule, answer_format, problem_hash, prompt_chars, metadata (JSON world state).
beam_traces: trace_id, problem_id, split, family, domain, difficulty, variant, model, reasoning_effort, prompt, reasoning, response, extracted_answer, answer, is_correct, format_ok, verified_correct, finish_reason, token counts, eval_set.
sft_verified: verified-correct Beam traces as chat messages with reasoning_content. preference_pairs: low vs high effort on the same problem, pair_type = correctness (one correct, one wrong) or concise_correct (both correct, chosen used at least 200 fewer tokens). model_evaluation: one row per model and evaluation item with the response, extracted answer, correctness, tokens, latency and failure category.
Limitations
- All problems are synthetic. They test careful rule-following and computation over realistic artefacts, not open-ended judgement, and the documents are simpler than real policies, ledgers or logs.
- Verification covers the key, not the reasoning text: a trace marked correct reached the right final answer, but its intermediate steps are not checked.
- Beam traces cover a stratified subset of the problems (812 problems), far fewer than the problem pool.
- The evaluation uses one sample per model and item through free hosted services; differences inside the confidence intervals should not be read as rankings. Some models completed fewer items because of rate limits and are reported separately with their own n.
- Difficulty levels are defined per family by input size and number of active rules, so the same level is not equally hard across families.
- Template-group splits hold out whole workplace skins, but the plain, structured and chained styles share instruction wording across splits.
License
Apache-2.0 for the problems, keys and metadata. Model outputs (Beam traces and evaluation responses) are included for research; check each model's own terms before reuse.
Citation
@misc{marshootlab2026beamverifiedv7,
title = {Beam Verified Reasoning v7: dual-verified applied reasoning problems, Beam traces and a held-out model evaluation},
author = {Marshoot Lab},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/MarshootLab/beam-verified-reasoning-traces}}
}
- Downloads last month
- 948