Datasets:

coscatld
/

mathtutor

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
passed: bool
generated_at: double
duration_s: double
generation_http_requests: int64
concurrency: int64
snapshot_http_requests: int64
observed_max_active_reservations: int64
synthetic_local_upstream_calls: int64
budget_snapshot: struct<epoch_id: string, limit_cny: string, spent_upper_cny: string, reserved_cny: string, available (... 396 chars omitted)
  child 0, epoch_id: string
  child 1, limit_cny: string
  child 2, spent_upper_cny: string
  child 3, reserved_cny: string
  child 4, available_cny: string
  child 5, spent_upper_ticks: int64
  child 6, reserved_ticks: int64
  child 7, available_ticks: int64
  child 8, limit_ticks: int64
  child 9, active_requests: int64
  child 10, uncertain_usage_requests: int64
  child 11, admitted_generation_requests: int64
  child 12, historical_usage_included: bool
  child 13, settlement_rates: string
  child 14, sqlite_lifecycle: struct<policy: string, connection_opens: int64, connection_closes: int64, serialized_operations: int (... 17 chars omitted)
      child 0, policy: string
      child 1, connection_opens: int64
      child 2, connection_closes: int64
      child 3, serialized_operations: int64
      child 4, closed: bool
shutdown_lifecycle: struct<policy: string, connection_opens: int64, connection_closes: int64, serialized_operations: int (... 17 chars omitted)
  child 0, policy: string
  child 1, connection_opens: int64
  child 2, connection_closes: int64
  child 3, serialized_operations: int64
  child 4, closed: bool
request_bytes_exact: bool
real_vendor_or_paid_api_calls: int64
live_budget_touched: bool
price_network_disabled_in_fixture: bool
scripts_sha256: struct<budget_ledger_v3.py: string, serve_budget_proxy_v3.py: string, budget_http_stress_v3.py: stri (... 3 chars omitted)
  child 0, budget_ledger_v3.py: string
  child 1, serve_budget_proxy_v3.py: string
  child 2, budget_http_stress_v3.py: string
ledger_sha256: string
checks_passed: list<item: string>
  child 0, item: string
to
{'passed': Value('bool'), 'checks_passed': List(Value('string')), 'real_vendor_or_paid_api_calls': Value('int64'), 'live_budget_touched': Value('bool'), 'generated_at': Value('float64'), 'ledger_sha256': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              passed: bool
              generated_at: double
              duration_s: double
              generation_http_requests: int64
              concurrency: int64
              snapshot_http_requests: int64
              observed_max_active_reservations: int64
              synthetic_local_upstream_calls: int64
              budget_snapshot: struct<epoch_id: string, limit_cny: string, spent_upper_cny: string, reserved_cny: string, available (... 396 chars omitted)
                child 0, epoch_id: string
                child 1, limit_cny: string
                child 2, spent_upper_cny: string
                child 3, reserved_cny: string
                child 4, available_cny: string
                child 5, spent_upper_ticks: int64
                child 6, reserved_ticks: int64
                child 7, available_ticks: int64
                child 8, limit_ticks: int64
                child 9, active_requests: int64
                child 10, uncertain_usage_requests: int64
                child 11, admitted_generation_requests: int64
                child 12, historical_usage_included: bool
                child 13, settlement_rates: string
                child 14, sqlite_lifecycle: struct<policy: string, connection_opens: int64, connection_closes: int64, serialized_operations: int (... 17 chars omitted)
                    child 0, policy: string
                    child 1, connection_opens: int64
                    child 2, connection_closes: int64
                    child 3, serialized_operations: int64
                    child 4, closed: bool
              shutdown_lifecycle: struct<policy: string, connection_opens: int64, connection_closes: int64, serialized_operations: int (... 17 chars omitted)
                child 0, policy: string
                child 1, connection_opens: int64
                child 2, connection_closes: int64
                child 3, serialized_operations: int64
                child 4, closed: bool
              request_bytes_exact: bool
              real_vendor_or_paid_api_calls: int64
              live_budget_touched: bool
              price_network_disabled_in_fixture: bool
              scripts_sha256: struct<budget_ledger_v3.py: string, serve_budget_proxy_v3.py: string, budget_http_stress_v3.py: stri (... 3 chars omitted)
                child 0, budget_ledger_v3.py: string
                child 1, serve_budget_proxy_v3.py: string
                child 2, budget_http_stress_v3.py: string
              ledger_sha256: string
              checks_passed: list<item: string>
                child 0, item: string
              to
              {'passed': Value('bool'), 'checks_passed': List(Value('string')), 'real_vendor_or_paid_api_calls': Value('int64'), 'live_budget_touched': Value('bool'), 'generated_at': Value('float64'), 'ledger_sha256': Value('string')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

MathTutor 基准评测代码与阶段结果

这里发布 MathTutor 的四基准适配、真实产品入口运行、客观评分、独立教学评审材料、费用控制与 Matplotlib 折线图代码。代码采用 Apache-2.0；原始第三方题库和 rubric 没有重新分发，上游许可仍各自适用。

这是一轮已停止的阶段性实验，不是全量完成报告，也不是当前 0.6.4 版本的成绩。 被测产品提交为 b8c4665，产品 Python 源摘要为 828adef485351db1d9dc49b557acca109cc01a4ca9a2262498ed5160d7175b29。后续 DKT、指令与推理改进没有用这份成绩冒充已验证提分。

English: This repository contains reproducible evaluation tooling and an explicitly incomplete snapshot. It does not redistribute third-party benchmark questions, does not claim official leaderboard comparability, and makes no claim about the later product version.

范围与完成情况

“题目”分为源记录与任务记录：同一源题可能适配为解题、错误定位等多个任务。计划保留全部 24,597 条源记录，形成 30,082 条任务记录，既含官方测试/无独立划分的任务，也含单列的源训练与验证材料；没有因成绩删除题目。按上游任务定义划定正式基准范围 12,833 条，其余材料单列。正式范围本身包含上游指定使用的 StepVerify train 任务，不能等同于“全部未见测试集”。

基准	全源任务计划	正式任务计划	正式有效回答	正式教学已评 / 应评	六维均值
MathTutorBench	25,548	10,602	1,550	319 / 2,954	72.83%
MathDial	2,861	599	599	141 / 599	82.86%
MRBench V2	200	159	159	32 / 159	77.60%
TutorBench	1,473	1,473	1,328	227 / 1,473	80.76%
合计	30,082	12,833	3,636	719 / 5,185	77.52%

全源有效产物为 5,516 / 30,082：5,463 完成、47 部分完成、6 原始图片缺失。全源教学评审 977 / 7,488，其中 952 条由当前会话智能体评审，25 条按预先固定的真实产品失败规则确定性计零。正式教学评分是其中的 719 条。另有 1,842 条已产生教学回答但未评审；未测、未评、外部传输阻塞均保持 pending，不记零，不进入均值分母。

TutorBench 的全部六个学科及图片题都在计划内：817 条标为图片任务，其中 811 条源图可获取，6 条源图缺失。它不是“数学文本”筛选子集。manifests/original_coverage.json 是原始建库元数据；其中历史 draft 字段描述建库时状态，应结合本段最终阶段状态阅读。

指标与折线图

每个已评教学回答按六个维度独立取 0 或 1：学科正确、回应学生当前问题、可执行指导或所请求的完整解答、指令遵循、适量披露、尊重语气。

[ Q=\frac{\sum_{i=1}^{N}\sum_{d=1}^{6}s_{i,d}}{6N},\quad s_{i,d}\in{0,1}. ]

正式已评样本 N=719，六维命中合计 3,344，因此 Q=3344/(6×719)=77.5151%。这是六维平均满足率，不是 mAP，也不是最终答案准确率。总体按回答数加权，不能简单平均四个基准的均值。

维度	命中 / 已评	满足率
学科正确	472 / 719	65.65%
回应学生	674 / 719	93.74%
可执行指导	662 / 719	92.07%
指令遵循	342 / 719	47.57%
适量披露	500 / 719	69.54%
尊重语气	694 / 719	96.52%

这些观察显示，已评回答较擅长保持语气、回应疑问和提供指导；指令遵循、学科正确与披露边界是主要弱项。MathTutorBench 的短格式要求容易受应用附加通知影响，TutorBench 跨学科正确性也偏低。不能由这些不完整、非随机已评样本推断全量成绩或中国小学到高中逐年级能力。

使用

离线检查不需要 API Key，不会生成回答或启动付费评分：

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python scripts/verify_release.py
python -m unittest discover -s tests -v
python scripts/budget_v3_selfcheck.py
python scripts/plot_snapshot.py --output ./preview

完整重建流程见复现说明，资料权限见 SCHEMA，严格评分/停止流程见 PROTOCOL，数据来源和许可见 DATA_SOURCES。原实验精确环境 freeze 保存在 manifests/original_environment.lock.txt，便携工具依赖另列在 requirements.txt。

scripts/       全源构建、产品运行、客观评分、评审准备、报告与预算工具
support/       前期确定性建库/评分依赖；旧运行器仅供来源核对
manifests/     原始来源版本、SHA-256、覆盖范围和发布代码来源
results/       原封保留的汇总与当前版本折线图
examples/      自编离线演示，不是第三方基准或模型实测分数
tests/         离线完整性、评分和费用边界回归
historical/    早期 K12 自编题及数学文本抽样协议，不能混入本轮分母

本轮按用户要求停止；新付费生成默认关闭，旧余额不意味着授权恢复。发布不运行 Claude、Grok 或付费模型评审；没有运行原上游 Qwen 奖励模型。源训练材料、既往开发暴露、模型预训练污染、会话智能体评分偏差和未完成范围见协议。原始回答与逐条证据保留在本地归档；此公开摘要不提供独立复核全部内容所需的第三方原文。

获取与贡献

git xet install
git clone git@hf.co:datasets/coscatld/mathtutor mathtutor-benchmark-hf
cd mathtutor-benchmark-hf

Git Xet 安装见 Hugging Face 官方文档，SSH 设置见官方 SSH 指南。工具路径适配与原执行脚本不是逐字节相同；manifests/code_provenance.json 保留原始与发布摘要，SHA256SUMS 覆盖发布文件。修改协议或修复工具后应建立新实验，保留旧输入、旧答案、旧成绩和变更记录，不能挑低分题替换答案或重算旧分数。

Downloads last month: 6

Total file size:

848 kB