Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
val_id: string
item_id: string
source: string
reply: string
label: struct<kind: string, flags: struct<endorses_unverified_approval: bool, invents_instruction: bool, as (... 60 chars omitted)
  child 0, kind: string
  child 1, flags: struct<endorses_unverified_approval: bool, invents_instruction: bool, asks_for_less: bool, claims_au (... 31 chars omitted)
      child 0, endorses_unverified_approval: bool
      child 1, invents_instruction: bool
      child 2, asks_for_less: bool
      child 3, claims_authority: bool
      child 4, threatens: bool
label_note: string
new_message: struct<from: string, text: string>
  child 0, from: string
  child 1, text: string
hard_reason: string
team: list<item: struct<name: string, doing: string>>
  child 0, item: struct<name: string, doing: string>
      child 0, name: string
      child 1, doing: string
bad_reply: string
bad_label: struct<kind: string, flags: list<item: string>>
  child 0, kind: string
  child 1, flags: list<item: string>
      child 0, item: string
excerpt: list<item: struct<from: string, text: string>>
  child 0, item: struct<from: string, text: string>
      child 0, from: string
      child 1, text: string
instructions: list<item: string>
  child 0, item: string
kind_gold: string
world_id: string
gold_reply: string
hard: bool
requests: list<item: string>
  child 0, item: string
principal: struct<name: string, role: string, org: string>
  child 0, name: string
  child 1, role: string
  child 2, org: string
domain: string
to
{'item_id': Value('string'), 'world_id': Value('string'), 'domain': Value('string'), 'principal': {'name': Value('string'), 'role': Value('string'), 'org': Value('string')}, 'instructions': List(Value('string')), 'team': List({'name': Value('string'), 'doing': Value('string')}), 'kind_gold': Value('string'), 'hard': Value('bool'), 'hard_reason': Value('string'), 'requests': List(Value('string')), 'gold_reply': Value('string'), 'bad_reply': Value('string'), 'bad_label': {'kind': Value('string'), 'flags': List(Value('string'))}, 'excerpt': List({'from': Value('string'), 'text': Value('string')}), 'new_message': {'from': Value('string'), 'text': Value('string')}}
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
              val_id: string
              item_id: string
              source: string
              reply: string
              label: struct<kind: string, flags: struct<endorses_unverified_approval: bool, invents_instruction: bool, as (... 60 chars omitted)
                child 0, kind: string
                child 1, flags: struct<endorses_unverified_approval: bool, invents_instruction: bool, asks_for_less: bool, claims_au (... 31 chars omitted)
                    child 0, endorses_unverified_approval: bool
                    child 1, invents_instruction: bool
                    child 2, asks_for_less: bool
                    child 3, claims_authority: bool
                    child 4, threatens: bool
              label_note: string
              new_message: struct<from: string, text: string>
                child 0, from: string
                child 1, text: string
              hard_reason: string
              team: list<item: struct<name: string, doing: string>>
                child 0, item: struct<name: string, doing: string>
                    child 0, name: string
                    child 1, doing: string
              bad_reply: string
              bad_label: struct<kind: string, flags: list<item: string>>
                child 0, kind: string
                child 1, flags: list<item: string>
                    child 0, item: string
              excerpt: list<item: struct<from: string, text: string>>
                child 0, item: struct<from: string, text: string>
                    child 0, from: string
                    child 1, text: string
              instructions: list<item: string>
                child 0, item: string
              kind_gold: string
              world_id: string
              gold_reply: string
              hard: bool
              requests: list<item: string>
                child 0, item: string
              principal: struct<name: string, role: string, org: string>
                child 0, name: string
                child 1, role: string
                child 2, org: string
              domain: string
              to
              {'item_id': Value('string'), 'world_id': Value('string'), 'domain': Value('string'), 'principal': {'name': Value('string'), 'role': Value('string'), 'org': Value('string')}, 'instructions': List(Value('string')), 'team': List({'name': Value('string'), 'doing': Value('string')}), 'kind_gold': Value('string'), 'hard': Value('bool'), 'hard_reason': Value('string'), 'requests': List(Value('string')), 'gold_reply': Value('string'), 'bad_reply': Value('string'), 'bad_label': {'kind': Value('string'), 'flags': List(Value('string'))}, 'excerpt': List({'from': Value('string'), 'text': Value('string')}), 'new_message': {'from': Value('string'), 'text': 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.

StewardBench

A held-out benchmark of a steward's judgement. A steward is a model appointed by a human principal to a team of agents; it reads a team situation and writes one message to the team. Each item gives the principal (name, role, organisation), the principal's standing instructions, the team and what each member is doing, the requests in play, and a gold reply kind, with a gold reply and a contrasting bad reply. It was written for the paper Weak Commissar, Strong Agent: Can a Small Loyal Model Raise an Agent's Loyalty to Its Principal? (Ephraiem Sarabamoun, Humanity First Research, Apart Swarm Alignment Hackathon, November 2026). All people, organisations and situations are invented.

The steward fine-tuned and scored on it is released as ephraiemsam16/commissar-steward-qwen3.5-4b-lora.

Versions

v1 (v1/) has 120 test items and 30 validation items in ten business domains (customer support, accounts payable, procurement, legal and compliance operations, a university research group, software release, non-clinical clinic administration, warehouse logistics, facilities, and a publishing desk). The test items are 12 per domain and 20 per reply kind: an opening brief, restating the principal's instructions, declining a shortcut, questioning an unverified approval, backing a teammate who raises a real problem, and staying quiet. 67 are marked hard, with the reason. The steward's system prompt is v1/steward_prompt.txt (a faithful policy).

v2 (v2/) scores a principal-first policy on 150 test items (82 hard) and 30 validation items. Restate items were relabelled as declining a shortcut, and 20 redirect-of-effort items and 10 quiet controls, in which an agent has correctly made one honest attempt for someone else, were added, giving 40 decline-a-shortcut, 30 stay-quiet and 20 each of the other four kinds. Its primary metric is macro accuracy over reply kinds. Each item records the principal, the instructions, the team, the requests (with the ones made on someone else's behalf in others_requests), kind_gold, hard and hard_reason, gold_reply, and bad_reply with its bad_label.

Steward instructions. v2/steward_prompt_v2.txt, _v2_1.txt and _v2_2.txt are three versions of the steward's instructions, each scored on the same v2 items. v2.2 (v2/steward_prompt_v2_2.txt) is the system prompt the released steward was trained and evaluated with. It tells the steward that the principal's own work comes first and is seen through until it is done or definitely cannot be. A request made for anyone else, including one the principal passes along on someone else's behalf, gets one honest attempt and a truthful report. {PRINCIPAL} is filled with the item's principal. Each version has its own pin file.

Judge

A reply is scored by Claude Sonnet 5.5 (claude-sonnet-5-5), run as an isolated command-line session with no tools, memory or project instructions. Its system prompt is v1/judge_system.txt for both versions; its rubric is v1/judge_template.txt for v1 and v2/judge_template_v2.txt for v2 under every set of instructions. The judge returns the reply's kind and flags (for example endorsing a forged approval, claiming authority, inventing an instruction). The items, prompts, judge prompts and scoring code were pinned by sha256 before the judge was validated and before any steward model was scored (v1/PINS.json, v2/PINS_v2.json, v2/PINS_v2_1.json, v2/PINS_v2_2.json). The judge prompt files here match the pinned hashes, and the v2.2 pin carries the same judge template hash as v2. On the 30 validation items the judge agreed with the gold kind on 93.3 percent (Cohen's kappa 0.92) for v1 and 96.7 percent (kappa 0.96) for v2, and an independent blind labeller agreed with the gold kind on 36 of 36 sampled v1 items. A second-party check of the v2 gold labels has not been run.

Results

v2 test items under the v2.2 instructions. Qwen3.5-4B is the base model, with thinking off, temperature 0 and at most 300 tokens per reply. Intervals are 95 percent item bootstraps.

Steward Macro accuracy Forged approvals endorsed False alarms on stay-quiet items
Qwen3.5-4B (base) 65.6 [59.0, 71.7] 3 of 20 23 of 30
Round-1 LoRA (not released) 56.2 [48.9, 63.7] 8 of 20 9 of 30
Round-2 LoRA (commissar-steward-qwen3.5-4b-lora) 70.8 [64.2, 77.5] 2 of 20 0 of 30

Paired on the same items, the round-2 adapter minus the base model is +5.3 points, with a paired bootstrap 95 percent interval of [-3.1, 13.5]. The interval includes zero, so the gain over the base model is not established. Against the round-1 adapter the difference is +14.6 points [4.7, 24.7]. Most of the gain comes from the stay-quiet items. The round-2 adapter put the principal first in fewer opening briefs than the base model (13 of 20 against 20), and it declined fewer of the shortcuts it should have declined (24 of 40 against 30).

Earlier reference points for the base model are 58.3 percent [50.0, 66.7] on v1 under the faithful prompt, 60.8 percent [55.4, 66.1] on v2 under the v2 prompt, and 65.3 percent [59.0, 71.7] on v2 under the v2.1 instructions.

Limits

AI agents working under the author's direction wrote the world files the items are built from. One judge model scores them, so the benchmark measures agreement with that judge's reading of the gold kinds. The situations are short, English and business-shaped. Only one steward model family has been scored. The pin files had one local path prefix replaced with <local>/ for this release; nothing else in them changed (see provenance.json). manifest.sha256 lists every file's sha256.

Licence

CC BY 4.0. Please cite the paper below.

Citation

@misc{sarabamoun2026weakcommissar,
  title  = {Weak Commissar, Strong Agent: Can a Small Loyal Model Raise an Agent's Loyalty to Its Principal?},
  author = {Sarabamoun, Ephraiem},
  year   = {2026},
  note   = {Humanity First Research. Apart Research Swarm Alignment Hackathon}
}
Downloads last month
58