Datasets:

HanyueShen
/

YunXiaoHe-RISA-Avalon-Eval

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
schema: string
note: string
runs: list<item: struct<dataset: string, repetition: int64, arm: string, receipt_sha256: string, calls: li (... 267 chars omitted)
  child 0, item: struct<dataset: string, repetition: int64, arm: string, receipt_sha256: string, calls: list<item: st (... 255 chars omitted)
      child 0, dataset: string
      child 1, repetition: int64
      child 2, arm: string
      child 3, receipt_sha256: string
      child 4, calls: list<item: struct<index: int64, purpose: string, provider: string, model: string, reported_fields: l (... 167 chars omitted)
          child 0, item: struct<index: int64, purpose: string, provider: string, model: string, reported_fields: list<item: s (... 155 chars omitted)
              child 0, index: int64
              child 1, purpose: string
              child 2, provider: string
              child 3, model: string
              child 4, reported_fields: list<item: string>
                  child 0, item: string
              child 5, usage: struct<inputTokens: int64, outputTokens: int64, cacheReadTokens: int64, cacheWriteTokens: null, reas (... 39 chars omitted)
                  child 0, inputTokens: int64
                  child 1, outputTokens: int64
                  child 2, cacheReadTokens: int64
                  child 3, cacheWriteTokens: null
                  child 4, reasoningTokens: int64
                  child 5, totalTokens: int64
provider: string
design: string
candidate: string
summary: list<item: struct<da
...
aseline_mean_quality: double
      child 9, candidate_mean_quality: double
      child 10, baseline_token_values: list<item: int64>
          child 0, item: int64
      child 11, candidate_token_values: list<item: int64>
          child 0, item: int64
      child 12, baseline_quality_values: list<item: double>
          child 0, item: double
      child 13, candidate_quality_values: list<item: double>
          child 0, item: double
profile: string
token_definition: string
model: string
optional_usage_fields: string
study_date: timestamp[s]
billing_cost_reduction_percent: null
limitations: list<item: string>
  child 0, item: string
task_reasoning_effort: string
optimization_policy: string
baseline: string
native_saved_model_replays: list<item: struct<dataset: string, repetition: int64, receipt: struct<model_sha256: string, rows: in (... 130 chars omitted)
  child 0, item: struct<dataset: string, repetition: int64, receipt: struct<model_sha256: string, rows: int64, provid (... 118 chars omitted)
      child 0, dataset: string
      child 1, repetition: int64
      child 2, receipt: struct<model_sha256: string, rows: int64, provider_calls: int64, training_calls: int64, max_abs_diff (... 65 chars omitted)
          child 0, model_sha256: string
          child 1, rows: int64
          child 2, provider_calls: int64
          child 3, training_calls: int64
          child 4, max_abs_difference_from_delivered_predictions: int64
          child 5, same_as_delivery: bool
sdk: string
to
{'schema': Value('string'), 'study_date': Value('timestamp[s]'), 'provider': Value('string'), 'model': Value('string'), 'task_reasoning_effort': Value('string'), 'sdk': Value('string'), 'profile': Value('string'), 'optimization_policy': Value('string'), 'candidate': Value('string'), 'baseline': Value('string'), 'design': Value('string'), 'token_definition': Value('string'), 'optional_usage_fields': Value('string'), 'billing_cost_reduction_percent': Value('null'), 'summary': List({'dataset': Value('string'), 'task': Value('string'), 'repetitions_per_arm': Value('int64'), 'heldout_rows': Value('int64'), 'baseline_mean_tokens': Value('float64'), 'candidate_mean_tokens': Value('float64'), 'token_reduction_percent': Value('float64'), 'quality_metric': Value('string'), 'baseline_mean_quality': Value('float64'), 'candidate_mean_quality': Value('float64'), 'baseline_token_values': List(Value('int64')), 'candidate_token_values': List(Value('int64')), 'baseline_quality_values': List(Value('float64')), 'candidate_quality_values': List(Value('float64'))}), 'runs': List({'dataset': Value('string'), 'task': Value('string'), 'repetition': Value('int64'), 'arm': Value('string'), 'evaluation_type': Value('string'), 'total_tokens': Value('int64'), 'input_tokens': Value('int64'), 'output_tokens': Value('int64'), 'cache_read_tokens': Value('int64'), 'reported_cache_write_tokens': Value('null'), 'reported_reasoning_tokens_included_in_output': Value('int64'), 'calls_without_reported_reasoning': Value('int64'), 'provider_requests': Value('int64'), 'elapsed_seconds': Value('float64'), 'heldout_rows': Value('int64'), 'accuracy': Value('float64'), 'rmse': Value('float64'), 'mae': Value('float64'), 'r2': Value('float64'), 'task_sha256': Value('string'), 'train_csv_sha256': Value('string'), 'predict_csv_sha256': Value('string'), 'prediction_sha256': Value('string'), 'usage_receipt_sha256': Value('string'), 'source_manifest_sha256': Value('string')}), 'native_saved_model_replays': List({'dataset': Value('string'), 'repetition': Value('int64'), 'receipt': {'model_sha256': Value('string'), 'rows': Value('int64'), 'provider_calls': Value('int64'), 'training_calls': Value('int64'), 'max_abs_difference_from_delivered_predictions': Value('int64'), 'same_as_delivery': Value('bool')}}), 'limitations': List(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
              schema: string
              note: string
              runs: list<item: struct<dataset: string, repetition: int64, arm: string, receipt_sha256: string, calls: li (... 267 chars omitted)
                child 0, item: struct<dataset: string, repetition: int64, arm: string, receipt_sha256: string, calls: list<item: st (... 255 chars omitted)
                    child 0, dataset: string
                    child 1, repetition: int64
                    child 2, arm: string
                    child 3, receipt_sha256: string
                    child 4, calls: list<item: struct<index: int64, purpose: string, provider: string, model: string, reported_fields: l (... 167 chars omitted)
                        child 0, item: struct<index: int64, purpose: string, provider: string, model: string, reported_fields: list<item: s (... 155 chars omitted)
                            child 0, index: int64
                            child 1, purpose: string
                            child 2, provider: string
                            child 3, model: string
                            child 4, reported_fields: list<item: string>
                                child 0, item: string
                            child 5, usage: struct<inputTokens: int64, outputTokens: int64, cacheReadTokens: int64, cacheWriteTokens: null, reas (... 39 chars omitted)
                                child 0, inputTokens: int64
                                child 1, outputTokens: int64
                                child 2, cacheReadTokens: int64
                                child 3, cacheWriteTokens: null
                                child 4, reasoningTokens: int64
                                child 5, totalTokens: int64
              provider: string
              design: string
              candidate: string
              summary: list<item: struct<da
              ...
              aseline_mean_quality: double
                    child 9, candidate_mean_quality: double
                    child 10, baseline_token_values: list<item: int64>
                        child 0, item: int64
                    child 11, candidate_token_values: list<item: int64>
                        child 0, item: int64
                    child 12, baseline_quality_values: list<item: double>
                        child 0, item: double
                    child 13, candidate_quality_values: list<item: double>
                        child 0, item: double
              profile: string
              token_definition: string
              model: string
              optional_usage_fields: string
              study_date: timestamp[s]
              billing_cost_reduction_percent: null
              limitations: list<item: string>
                child 0, item: string
              task_reasoning_effort: string
              optimization_policy: string
              baseline: string
              native_saved_model_replays: list<item: struct<dataset: string, repetition: int64, receipt: struct<model_sha256: string, rows: in (... 130 chars omitted)
                child 0, item: struct<dataset: string, repetition: int64, receipt: struct<model_sha256: string, rows: int64, provid (... 118 chars omitted)
                    child 0, dataset: string
                    child 1, repetition: int64
                    child 2, receipt: struct<model_sha256: string, rows: int64, provider_calls: int64, training_calls: int64, max_abs_diff (... 65 chars omitted)
                        child 0, model_sha256: string
                        child 1, rows: int64
                        child 2, provider_calls: int64
                        child 3, training_calls: int64
                        child 4, max_abs_difference_from_delivered_predictions: int64
                        child 5, same_as_delivery: bool
              sdk: string
              to
              {'schema': Value('string'), 'study_date': Value('timestamp[s]'), 'provider': Value('string'), 'model': Value('string'), 'task_reasoning_effort': Value('string'), 'sdk': Value('string'), 'profile': Value('string'), 'optimization_policy': Value('string'), 'candidate': Value('string'), 'baseline': Value('string'), 'design': Value('string'), 'token_definition': Value('string'), 'optional_usage_fields': Value('string'), 'billing_cost_reduction_percent': Value('null'), 'summary': List({'dataset': Value('string'), 'task': Value('string'), 'repetitions_per_arm': Value('int64'), 'heldout_rows': Value('int64'), 'baseline_mean_tokens': Value('float64'), 'candidate_mean_tokens': Value('float64'), 'token_reduction_percent': Value('float64'), 'quality_metric': Value('string'), 'baseline_mean_quality': Value('float64'), 'candidate_mean_quality': Value('float64'), 'baseline_token_values': List(Value('int64')), 'candidate_token_values': List(Value('int64')), 'baseline_quality_values': List(Value('float64')), 'candidate_quality_values': List(Value('float64'))}), 'runs': List({'dataset': Value('string'), 'task': Value('string'), 'repetition': Value('int64'), 'arm': Value('string'), 'evaluation_type': Value('string'), 'total_tokens': Value('int64'), 'input_tokens': Value('int64'), 'output_tokens': Value('int64'), 'cache_read_tokens': Value('int64'), 'reported_cache_write_tokens': Value('null'), 'reported_reasoning_tokens_included_in_output': Value('int64'), 'calls_without_reported_reasoning': Value('int64'), 'provider_requests': Value('int64'), 'elapsed_seconds': Value('float64'), 'heldout_rows': Value('int64'), 'accuracy': Value('float64'), 'rmse': Value('float64'), 'mae': Value('float64'), 'r2': Value('float64'), 'task_sha256': Value('string'), 'train_csv_sha256': Value('string'), 'predict_csv_sha256': Value('string'), 'prediction_sha256': Value('string'), 'usage_receipt_sha256': Value('string'), 'source_manifest_sha256': Value('string')}), 'native_saved_model_replays': List({'dataset': Value('string'), 'repetition': Value('int64'), 'receipt': {'model_sha256': Value('string'), 'rows': Value('int64'), 'provider_calls': Value('int64'), 'training_calls': Value('int64'), 'max_abs_difference_from_delivered_predictions': Value('int64'), 'same_as_delivery': Value('bool')}}), 'limitations': List(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.

云小鹤 0.3.9.8：RISA 与 Avalon 公开评测

这份成果集记录云小鹤在两类公开任务中的完整成绩：跨数据集表格建模，以及有状态、长时程的供应链决策。这里可以看到版本、能力、评测口径、结果、图和证据哈希；云小鹤的私有工作系统、RISA 与 Avalon 的实现没有公开。

2026-10-05 更新：完整 30 任务复现与 token 效率

新增完整中英报告与逐次数据，以原 TabArena 完整 30 个二分类任务的复现为主要建模验证：宏平均 ROC-AUC **81.669% → 81.669%，原质量保留 100%**。独立五任务同模型对照的 token 减少 62.7%、准确率保留 98.9% 单独呈现，不将其当作 Full-30 的新用量测量。

r7 发布前复现，2026-10-03：保存的应用与原始 SDK 在 DGX ARM64 上重跑原 TabArena 二分类 30 任务、472 候选，指标保持一致，全部最终预测排名不变。原宏平均 ROC-AUC 0.816690 因相同标签下排名一致而保留；不是新评分，也没有新的 DSH token 测量。最终下载包的安装验收与该次 Full-30 复现分开记录。
补充 DSH 接入记录，2026-10-02：报告附有普通 DSH 与接入 RISA 的历史三任务对照、逐次用量与质量取舍，不将其作为 Full-30 的计量结果。

English: The bilingual October report leads with the original 30-task TabArena reproduction: macro ROC-AUC 81.669% → 81.669%, with 100% of mean quality retained through same-label rank equivalence. The separate historical five-task same-model study reports 62.7% fewer tokens and 98.9% mean accuracy retained; those token savings are not attributed to Full-30. Real DSH integration tests, including their regression-quality tradeoffs, remain available as an additional record. The September results below are unchanged.

一眼看懂结果

评测	结果	说明
TabArena 30 数据集	macro ROC-AUC 0.816690	较既有同集结果 +3.60 个百分点，24/30 数据集胜出
五任务同模型对照	Token -62.7%	accuracy 保留 98.9%
SupplyChainBench	54.810907	16/16 局、576/576 动作全部完成
冻结公开榜位置	完整模型点估计第 1	若加入冻结快照；较 Muse Spark 1.2 分数高 6.7%

使用的云小鹤能力

云小鹤版本：0.3.9.8
主模型：DeepSeek Pro，non-fast
云小鹤 - 数据治理Agent
RISA数据能力交换和操作协议：G1、G2、T1、注册特征工程与原生五折
云小鹤 - 供应链分析Agent（Avalon）r4
完成态校验、环境回执、失败重试、分项 Token 计量与独立审计

这些能力在同一项任务里协同工作。用户面对的是一个连续的首席智能体，而不是几个互不相干的服务。

两组结果说明了什么

RISA 路线证明云小鹤能把数据检查、特征处理、候选选择、五折验证和冻结评分连成一条可审查的建模过程。30 数据集结果提供广度与质量证据；五任务同模型对照则展示资源效率：Token 下降超过六成，accuracy 仍保留 98.9%。

Avalon 路线检验的是连续行动。系统不是写一份供应链建议，而是在冻结环境中完成 576 个原生动作，每一步都由环境执行并返回新状态。若将结果加入冻结公开榜，云小鹤在完整覆盖模型条目中点估计排第 1。

五任务同模型对照是独立的历史效率消融，用来展示同一模型接入 RISA 前后的 Token—质量变化，不与 Full-30 质量主结果混作同一次实验。

文件导航

reports/2A_云小鹤_RISA_Avalon_公开技术报告_2026-09-06.pdf：论文式技术报告
reports/2B_云小鹤_RISA_Avalon_客户成果报告_2026-09-06.pdf：面向客户与决策者的成果简报
data/benchmark_summary_public.json：机器可读汇总
data/risa_full30_scores.csv：30 个数据集的配对结果
data/risa_five_task_token_ablation.csv：历史五任务同模型 Token—质量对照
data/avalon_leaderboard.csv：冻结公开榜快照与本轮结果
evidence/evidence_hashes.json：关键证据哈希
figures/：PNG、SVG 与 PDF 三种图形格式

评测口径

RISA 的 0.816690 对应 TabArena v0.1 binary r0f0s0 公共 30 数据集子集，不代表完整 51 数据集总榜。Avalon 使用 SupplyChainBench 的冻结 16 个种子、相同环境与成本公式；“第 1”表示若加入冻结快照后，在完整覆盖模型条目中的点估计位置。环境成本是 benchmark 单位，不是美元。

上游项目：

TabArena：https://github.com/autogluon/tabarena
SupplyChainBench：https://github.com/ConstantinVictorBeatErtel/Supply_Chain_Bench
冻结公开评测目录：https://github.com/ConstantinVictorBeatErtel/Supply_Chain_Bench/tree/main/artifacts/live_y_domain_randomized_grpo_v2/evaluations

权利与复用

公开报告、派生汇总与图用于结果核验和交流。上游 benchmark、数据与模型仍遵循各自许可。本仓库不授予云小鹤私有工作系统、RISA 或 Avalon 实现的源码许可，也不包含任何 API 凭据。

Downloads last month: 810

Total file size:

4.38 MB