The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
share_search: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
share_page: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
share_official: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
share_data: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
share_process: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
domains: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
moves:
...
<task: string, model: string, repeat: string, acc: double, moves: int64, share_search: double, (... 143 chars omitted)
child 0, task: string
child 1, model: string
child 2, repeat: string
child 3, acc: double
child 4, moves: int64
child 5, share_search: double
child 6, share_page: double
child 7, share_official: double
child 8, share_data: double
child 9, share_process: double
child 10, domains: int64
child 11, reads_per_search: double
child 12, wall: double
T08-GLM-r2: struct<task: string, model: string, repeat: string, acc: double, moves: int64, share_search: double, (... 143 chars omitted)
child 0, task: string
child 1, model: string
child 2, repeat: string
child 3, acc: double
child 4, moves: int64
child 5, share_search: double
child 6, share_page: double
child 7, share_official: double
child 8, share_data: double
child 9, share_process: double
child 10, domains: int64
child 11, reads_per_search: double
child 12, wall: double
T06-GLM-r4: struct<task: string, model: string, repeat: string, acc: double, moves: int64, share_search: double, (... 141 chars omitted)
child 0, task: string
child 1, model: string
child 2, repeat: string
child 3, acc: double
child 4, moves: int64
child 5, share_search: double
child 6, share_page: double
child 7, share_official: double
child 8, share_data: double
child 9, share_process: double
child 10, domains: int64
child 11, reads_per_search: null
child 12, wall: double
to
{'T01-DS-r1': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('float64'), 'wall': Value('float64')}, 'T01-DS-r2': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('float64'), 'wall': Value('float64')}, 'T01-DS-r3': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('float64'), 'wall': Value('float64')}, 'T01-DS-r4': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process
...
Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('null'), 'wall': Value('float64')}, 'T12-P1-r3': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('null'), 'wall': Value('float64')}, 'T12-P1-r4': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('null'), 'wall': Value('float64')}, 'T12-P1-r5': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('null'), 'wall': Value('float64')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, 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 129, 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 544, 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 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, 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 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, 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 361, 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 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
share_search: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
share_page: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
share_official: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
share_data: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
share_process: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
domains: struct<byModel: struct<P1: double, DS: double, GLM: double>, p: double>
child 0, byModel: struct<P1: double, DS: double, GLM: double>
child 0, P1: double
child 1, DS: double
child 2, GLM: double
child 1, p: double
moves:
...
<task: string, model: string, repeat: string, acc: double, moves: int64, share_search: double, (... 143 chars omitted)
child 0, task: string
child 1, model: string
child 2, repeat: string
child 3, acc: double
child 4, moves: int64
child 5, share_search: double
child 6, share_page: double
child 7, share_official: double
child 8, share_data: double
child 9, share_process: double
child 10, domains: int64
child 11, reads_per_search: double
child 12, wall: double
T08-GLM-r2: struct<task: string, model: string, repeat: string, acc: double, moves: int64, share_search: double, (... 143 chars omitted)
child 0, task: string
child 1, model: string
child 2, repeat: string
child 3, acc: double
child 4, moves: int64
child 5, share_search: double
child 6, share_page: double
child 7, share_official: double
child 8, share_data: double
child 9, share_process: double
child 10, domains: int64
child 11, reads_per_search: double
child 12, wall: double
T06-GLM-r4: struct<task: string, model: string, repeat: string, acc: double, moves: int64, share_search: double, (... 141 chars omitted)
child 0, task: string
child 1, model: string
child 2, repeat: string
child 3, acc: double
child 4, moves: int64
child 5, share_search: double
child 6, share_page: double
child 7, share_official: double
child 8, share_data: double
child 9, share_process: double
child 10, domains: int64
child 11, reads_per_search: null
child 12, wall: double
to
{'T01-DS-r1': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('float64'), 'wall': Value('float64')}, 'T01-DS-r2': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('float64'), 'wall': Value('float64')}, 'T01-DS-r3': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('float64'), 'wall': Value('float64')}, 'T01-DS-r4': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process
...
Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('null'), 'wall': Value('float64')}, 'T12-P1-r3': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('null'), 'wall': Value('float64')}, 'T12-P1-r4': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('null'), 'wall': Value('float64')}, 'T12-P1-r5': {'task': Value('string'), 'model': Value('string'), 'repeat': Value('string'), 'acc': Value('float64'), 'moves': Value('int64'), 'share_search': Value('float64'), 'share_page': Value('float64'), 'share_official': Value('float64'), 'share_data': Value('float64'), 'share_process': Value('float64'), 'domains': Value('int64'), 'reads_per_search': Value('null'), 'wall': Value('float64')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Particle-1, DeepSeek V4.1 Flash and GLM-5.3-Flash in Claude Code with hedwigai workbooks
We tested three models as the model behind Claude Code, doing research in hedwigai workbooks: Particle-1 (particle-ai/particle-1, a DeepSeek model with further post-training), DeepSeek V4.1 Flash (deepseek/deepseek-v4.1-flash-1017) and GLM-5.3-Flash (zai-org/GLM-5.3-Flash). Only the model changed. Each model did 12 tasks 5 times, 180 runs in all. Every delivered document was graded blind against an answer key written before the runs.
Each run worked in a fresh workbook seeded with the task's tabs. It read and edited tabs as files, with a shell, web search and web fetch, and no other data. Every write is a version that records the sources it cites, which is what makes citation precision measurable.
Write-up: https://www.hedwigai.com/blog/model-bench
What is in it
| path | contents |
|---|---|
runs.jsonl |
One row per run. A row holds the task, model, repeat and step; the deployed revisions at the start and end of the run; wall time, model calls, prompt, cached and completion tokens, turns and the stop reason; the collection checks; and the blind grade (score, plus score_grader_B where a second grader graded it). |
documents/<run>.md |
The document as delivered, every tab. |
traces/<run>.jsonl |
The run's event stream: harness output lines, model-call start and return with token counts, tool returns with duration and status, and drafts. Every string field is cut to 2 KB. |
answer_key/ |
tasks.md (seed tabs and instructions, exactly as given) and T01.json … T12.json. Each fact in a key carries the URLs that confirm it. |
rubric.md, grade_schema.json |
The grading rules and the shape of a grade. |
grades/A/, grades/B/ |
The grades themselves, one JSON file per graded document, named by blind id. |
grader_agreement.json |
Cohen's kappa between grader A and grader B, and how it was computed. |
A run is named T<NN>-<model>-r<N>. T12 is two runs on one workbook, so it adds -step1 and -step2, and only step 2 carries a grade. Model codes are P1 (particle-1), DS (deepseek-v4.1-flash-1017) and GLM (GLM-5.3-Flash).
The tasks
There are 12 task types, each repeated with all three models.
- Latest primary funding round for 6 private companies. Tender offers and secondaries are a trap.
- FDA novel approvals in one month.
- Form D biotech raises by filing date.
- 10-K revenue and net income for named fiscal years.
- Contract terms as amended, read from a tab in the workbook.
- Currency conversion at an ECB reference rate, with a total.
- US importers of black tea (open-ended).
- Updating 2 stale rows of 10 while leaving the rest byte-identical.
- Formula column and totals in an xlsx.
- HQ city and current CEO for 15 companies.
- Questions with no answer, where saying so is correct and fabrication is the failure.
- A follow-up instruction on the first run's output.
How runs were collected
- Environment. Each run worked in its own workbook. It could read its seed tabs, search and fetch the web, and use a shell, and could read no other data on the account. Each trace was checked for this (
checks.sealed_ok) and for the model it used (checks.model_ok). - Time limit. 15 minutes a run; T10, a breadth task, 30.
- Request settings. Every model ran with the same defaults: high reasoning,
top_p0.85 and a 32k completion cap. A run that filled its context compacted it and continued. - Outcomes. Timeouts and refusals are results. Their document is whatever stood when the run ended. A run stopped by infrastructure before it finished was replaced by a fresh run; only the replacement is included (
replaces_interrupted_or_rev_changed_run: true).
How runs were graded
Documents were anonymised before grading: model names, workbook and run ids, timings and token counts were removed. Grader A graded every document. Grader B, a different model, independently graded a seeded, task-stratified 20% (35 documents).
Each grade records:
- rows correct against the answer key, and accuracy;
- fabricated facts, meaning values found in no source;
- wrong citations, meaning the cited page was fetched and does not support the claim;
- citations that could not be fetched (excluded from precision rather than guessed);
- silent omissions, internal inconsistencies and instruction deviations.
rubric.md defines each count precisely.
Results
Accuracy is the share of a task's rows that match the key (a partly right row counts half, a run that delivered nothing counts 0), averaged over runs, with 95% bootstrap intervals (10,000 resamples of runs); 60 graded runs per model.
| model | rows right (95% CI) | runs with a made-up value | citation precision (of fetched) | median wall s |
|---|---|---|---|---|
| particle-1 | 91.2% (84.3–96.9%) | 8.3% | 96.3% | 375 |
| deepseek-v4.1-flash-1017 | 89.4% (82.2–95.8%) | 6.7% | 95.1% | 315 |
| GLM-5.3-Flash | 90.8% (84.2–96.5%) | 6.7% | 94.7% | 144 |
- Accuracy and citations. Equal within error: the intervals overlap. Citation precision is the share of fetched citations whose page supports the row: 1 − wrong / (checked − unfetchable). In 6 of the 12 tasks every run of every model got every row.
- Sources reached. T10 asks for each company's CEO as of 2026-11-12; NVIDIA's had resigned the day before. Particle-1 reported the change in 5 of 5 runs, DeepSeek V4.1 Flash in 3 of 5, GLM-5.3-Flash in 0 of 5. Particle-1 against GLM-5.3-Flash is significant (Fisher's exact test, p = 0.008); against DeepSeek V4.1 Flash it is not (p = 0.44).
- Time and tokens. GLM-5.3-Flash's median run used 144 s and 7,752 output tokens, against 375 s and 13,988 for Particle-1.
- Grader agreement. Grader A and grader B agreed on every row of the 35 double-graded documents (Cohen's kappa 1.0). Most rows are easy, so read that as a floor on grading noise, not proof of agreement on hard cases.
- Footnote. Two DeepSeek runs (T05 r1, T09 r2) ended on a harness error rather than a model one and count as 0 above; without them DeepSeek gets 92.5% of rows right.
Behaviour: what separated the runs
The traces support a behavioural reading of the results. analysis/ holds the scripts and their outputs. Every measure is a rule over trace lines, not a judgment by a model.
- Same research, different timing. A lookup is a web search, a page read, or a call to a data tool (SEC filings, trade, …). Lookups per run were alike (Particle-1 26.6, DeepSeek 24.2, GLM 25.5; permutation test p = 0.86), and so was the mix of kinds (every share p ≥ 0.19). Lookups after the run's first write to its draft were not: 14.0, 11.4 and 3.4 (p < 0.001). 45%, 47% and 20% of runs looked anything up after writing (p = 0.001). Permutation tests shuffle model labels within each task and repeat.
- T10 routes. Every traced run first took NVIDIA's CEO from the February 10-K (Huang). Three Particle-1 runs then pulled every company's latest 8-K and found Kress. Five runs (Particle-1 and DeepSeek) saw a search result hinting at a change and followed it to NVIDIA's release. One DeepSeek run saw such a result and did not follow it. No GLM run looked beyond the periodic reports; two read the 10-Q, which still lists Huang.
analysis/NOTES.mdgives the trace lines for each run. DeepSeek T10 r4 has no trace. - Queries carry beliefs. 123 of 128 T10 web searches about a CEO named a person, as a rule the expected incumbent (all three models).
- Proxy, not fabrication. Three of five GLM T10 documents state that the CEO column comes from the latest SEC filing; one states that the filings predate the question date.
- Checking and accuracy. Within the same task, runs that looked something up after writing scored 5.9 points higher (permutation within tasks, p = 0.07): suggestive, correlational.
Use the same workbooks from your own assistant
The environment is not specific to this benchmark. Any harness or assistant that takes a remote MCP server can create, read and write hedwigai workbooks and start research runs on them, at https://mcp.hedwigai.com. In Claude Code:
claude mcp add --transport http hedwigai https://mcp.hedwigai.com
then /mcp to sign in. Claude Desktop and claude.ai add it from https://claude.ai/directory/hedwigai; Grok, ChatGPT, Cursor and Codex connect to the same address (steps at https://www.hedwigai.com/agents).
Privacy
All runs were made on one internal account. The export replaces workbook ids, run ids, run directory paths, email addresses, bearer and API tokens, token query parameters and account ids with placeholders such as [workbook], [run] and [email]. A residual scan checks that none remain. Public email addresses quoted in fetched web pages are also replaced.
Limits
- Small sample. There are a few repeats per task per model, so read the bootstrap intervals rather than the means alone.
- As-of facts. The answer keys were written on 2026-11-12; facts after that date are out of scope. The NVIDIA CEO entry was corrected to Colette M. Kress (interim) from NVIDIA's 8-K of that day, and that row re-scored for every run (
answer_key/KEY_CHANGES.md). - Revisions. The harness and tools are deployed services;
revs_startandrevs_endrecord the revisions each run used. - T09 formulas. Blind graders judged T09 from the markdown table, which shows formula strings. They did not check the downloaded xlsx.
- T07 sources. The bill-of-lading site most T07 rows cite refuses automated fetches, so most T07 rows are graded
unverifiableand counted as not correct; T07 is low for all three models for that reason.
- Downloads last month
- 97