The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: TypeError
Message: Couldn't cast array of type
struct<NormalMessageResponded: int64>
to
{'pre_llm_call': Value('int64'), 'pre_tool_call': Value('int64')}
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 2312, in cast_table_to_schema
cast_array_to_feature(
~~~~~~~~~~~~~~~~~~~~~^
table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
feature,
^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2158, in cast_array_to_feature
raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
TypeError: Couldn't cast array of type
struct<NormalMessageResponded: int64>
to
{'pre_llm_call': Value('int64'), 'pre_tool_call': Value('int64')}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.
iLang Judge Observations
Every decision an agent was about to make, judged by the iLang v5.0 reference function, one line per event, published daily by the agent's operator. The text of the conversation never leaves the operator's machine. What is published is what the judgment layer saw and decided: the eleven-dimension vector, its probabilities and confidences, the mode f_v5 computed, the deciding step, the five-class answer the same model gave when asked directly in the same call, and whether that answer disagreed with the mode f_v5 computed from its own dimensions.
The judge is the open ilang-judge endpoint: eleven three-level rubric questions (one per dimension, anchored on the canon's own T:0.00 and T:1.00 definitions; from rubric v1 on, four of them carry one clarification each, written from the judge's error ledger) plus one five-class question, asked of TypeSafe's Jev in one call; the expected value of each dimension becomes the vector; f_v5 from the sealed canon (ilang-spec cad65e2, v5.0-pre-2.4.1-sealed) maps the vector to a mode; code writes the reason. The endpoint is stateless and stores nothing; the operator's bots write the returned record into their own logs, and the exporter in the same repository turns those logs into the files here.
Files
| Path | Contents |
|---|---|
data/<caller>/<day>.jsonl |
one observation per line for that caller and business day (Asia/Shanghai) |
daily/<caller>/<day>.json |
the day's counts: records, hooks, mode and class distributions, disagreement rate, latency percentiles, sessions |
Callers so far: feishu (a Feishu group bot for a paid course community, built on Hermes) and wecom (a WeChat Work customer-service bot, built on LangBot). Both belong to the same operator. A day's file is replaced when the day is re-exported, so the most recent day may be partial until the next run.
One observation
| field | meaning |
|---|---|
schema |
schema version, 1 |
caller, hook |
which bot, and where in its loop: pre_llm_call (before a reply is generated), pre_tool_call (before a tool runs), NormalMessageResponded (before an outgoing reply is sent) |
ts, day |
UTC second of the judgment; business day in Asia/Shanghai |
session, sender |
HMAC pseudonyms (16 hex) of the bot's session and sender ids, computed with a salt that stays on the operator's machine: events of one session are linkable, the ids are not recoverable |
tool_name |
for pre_tool_call, the tool's name; arguments are never exported |
status, http, roundtrip_ms, error |
whether the judge answered, and how long the bot's background call took (the bot never waits for it) |
id, mode, class5_key, reason |
the judge's record id, the f_v5 mode (M1 to M8), its five-class bucket (act, confirm, ask, hand_over, decline_or_stop), and the deciding step written by code, such as STEP-2 cer=0.24<0.30 |
vector, probabilities, confidence |
the eleven dimensions (two decimals), the three-level probabilities each came from, and the model's confidence per dimension |
jev_class5_key, jev_class5_probabilities, disagree |
the model's direct five-class answer (question twelve of the same call), its distribution, and whether it differs from the five-class bucket of the f_v5 mode. This is an internal-consistency measure between the model's direct judgment and the canon's function over the model's own perception; it is not a comparison with an independent judge or a prose baseline. A mode of M5 never counts as disagreement |
latency_ms, usage, versions |
the model call's latency and token usage; canon commit, judge core, rubric and Jev model versions |
Nothing else is exported. The exporter copies a whitelist of fields; the only free-text field, reason, is accepted only when it matches the shape the judge generates from numbers; identifiers are pseudonymised; text summaries, text hashes and error details are dropped. Each export runs a leak check against its own input: every private string of the log must be absent from the output, or nothing is uploaded.
Why publish this
The judgment layer of iLang v5.0 claims that a fixed function over eleven perceived dimensions can say what an agent should do next. Whether that holds is an empirical question, and the answer needs many decisions from real operation, with the operator's later corrections as the ground truth. This dataset is the first such stream. The operator's corrections are not in it yet; the first three-day analysis (2026-09-29) will be published beside it, and the same exporter lets any other operator of the open judge publish a comparable stream.
Provenance and terms
Observations start 2026-09-26 16:42 Asia/Shanghai. Operator: iLang Inc. (Long Quan Zhu). Judge core 0.1.0 until 2026-09-30 13:32 Asia/Shanghai and 0.2.0 until 14:55, both with rubric v0; core 0.3.0 with rubric v4 until 15:01; core 0.3.1 with rubric v5 from then on; every line records its own versions; the rubric versions and why each was made are in the repository's README. Canon ilang-spec cad65e2, perception model TypeSafe jev-latest (recorded per line). License CC BY 4.0. The conversation texts behind these observations belong to the people who wrote them and are not part of this dataset.
- Downloads last month
- 572