Hugging Face's logo

Datasets:
moebiusT7
/
l0-essentials-compact-ablation

Tasks:
Text Generation
Languages:
English
Japanese
Tags:
gemma4
gpt-oss
qwen3
deepseek
governance
answer-entitlement
License:
Dataset card Files
xet
 Community
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
variant: string
probe: string
seed: int64
score: double
detail: string
wall: double
tokens: int64
think_chars: int64
content: string
tool_calls: list<item: struct<type: string, function: struct<name: string, arguments: string>, id: string>>
  child 0, item: struct<type: string, function: struct<name: string, arguments: string>, id: string>
      child 0, type: string
      child 1, function: struct<name: string, arguments: string>
          child 0, name: string
          child 1, arguments: string
      child 2, id: string
policy: string
premise_validity: struct<name: string, rule: string, examples: list<item: string>, fail_action: string>
  child 0, name: string
  child 1, rule: string
  child 2, examples: list<item: string>
      child 0, item: string
  child 3, fail_action: string
routes: struct<ask: struct<trigger: string, behavior: string>, verify: struct<trigger: string, behavior: str (... 111 chars omitted)
  child 0, ask: struct<trigger: string, behavior: string>
      child 0, trigger: string
      child 1, behavior: string
  child 1, verify: struct<trigger: string, behavior: string>
      child 0, trigger: string
      child 1, behavior: string
  child 2, re_anchor: struct<trigger: string, behavior: string>
      child 0, trigger: string
      child 1, behavior: string
  child 3, abstain: struct<trigger: string, behavior: string>
      child 0, trigger: string
      child 1, behavior: string
to
{'policy': Value('string'), 'routes': {'ask': {'trigger': Value('string'), 'behavior': Value('string')}, 'verify': {'trigger': Value('string'), 'behavior': Value('string')}, 're_anchor': {'trigger': Value('string'), 'behavior': Value('string')}, 'abstain': {'trigger': Value('string'), 'behavior': Value('string')}}, 'premise_validity': {'name': Value('string'), 'rule': Value('string'), 'examples': List(Value('string')), 'fail_action': 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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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
              variant: string
              probe: string
              seed: int64
              score: double
              detail: string
              wall: double
              tokens: int64
              think_chars: int64
              content: string
              tool_calls: list<item: struct<type: string, function: struct<name: string, arguments: string>, id: string>>
                child 0, item: struct<type: string, function: struct<name: string, arguments: string>, id: string>
                    child 0, type: string
                    child 1, function: struct<name: string, arguments: string>
                        child 0, name: string
                        child 1, arguments: string
                    child 2, id: string
              policy: string
              premise_validity: struct<name: string, rule: string, examples: list<item: string>, fail_action: string>
                child 0, name: string
                child 1, rule: string
                child 2, examples: list<item: string>
                    child 0, item: string
                child 3, fail_action: string
              routes: struct<ask: struct<trigger: string, behavior: string>, verify: struct<trigger: string, behavior: str (... 111 chars omitted)
                child 0, ask: struct<trigger: string, behavior: string>
                    child 0, trigger: string
                    child 1, behavior: string
                child 1, verify: struct<trigger: string, behavior: string>
                    child 0, trigger: string
                    child 1, behavior: string
                child 2, re_anchor: struct<trigger: string, behavior: string>
                    child 0, trigger: string
                    child 1, behavior: string
                child 3, abstain: struct<trigger: string, behavior: string>
                    child 0, trigger: string
                    child 1, behavior: string
              to
              {'policy': Value('string'), 'routes': {'ask': {'trigger': Value('string'), 'behavior': Value('string')}, 'verify': {'trigger': Value('string'), 'behavior': Value('string')}, 're_anchor': {'trigger': Value('string'), 'behavior': Value('string')}, 'abstain': {'trigger': Value('string'), 'behavior': Value('string')}}, 'premise_validity': {'name': Value('string'), 'rule': Value('string'), 'examples': List(Value('string')), 'fail_action': 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.

L0 Essentials compact ablation — raw rows

Row-level outputs from ablating the MOBIUS MMV L0 Essentials governance prompt on six local models, three axes (tool loop / false premise / abstain chat), with the scorer sources and the pre-registered predictions.

Status: experimental; not adversarially reviewed. The scorers had five documented defects during the work (all fixed, rows rescored); raw outputs are included so you can rescore with your own instrument.

Headline numbers (see the validation note for full tables)

  • Tool loop: a ~480-token subset of L0 Essentials makes five local models ask before an irreversible tool instead of firing it (qwen3:30b-a3b: sent the email 3/3 → 0/3), at 1.0–2.3× tokens/step vs 1.0–5.5× for the full document.
  • False premise: on the two models that fabricate (gpt-oss-20b, deepseek-r1:14b) the subset cuts fabrication from 6–9/12 to 0–2/12; the full document leaves 4–5.
  • Personal/high-stakes chat: the subset with v8.4.1 abstain wording reaches 9/9 "decline + general information" on gemma4:26b, equal to the full document.
  • Exception: Qwen3.8-27B prefers the full document on the false-premise axis.
  • Adding any single dropped section back to the subset changes nothing on the chat axis (15 arms, all 9/9 on gemma4:26b).
  • Full document side effect (qwen3:30b-a3b refusing an ordinary file read 2/3) does not occur under the subset.

Predictions were written before every run; 24 of 38 were wrong. The scorers were fixed six times on real rows. This dataset exists so that the rows, not the narrative, are the artifact.

Files

path rows what
agent_ablation/gemma4-26b_A-F.json, …_F.json 108 six system layers × six tool-loop probes × 3 seeds
agent_transfer/gpt-oss-20b.json, qwen3-30b-a3b.json 24 each four layers × two decisive probes
premise/{gemma26,gptoss,deepseek}.json 48 each four layers × four false-premise questions
compact_v1/A_*.json, B_*.json — compact v1 and the rejected plain rewrite, both axes
compact_v2_rejected/A_*.json, B_*.json — v2 vs v1 on five models
compact_v1_1/A_*.json, B_*.json, C_*.json — v1.1 vs v1, three axes; C_ is the chat-abstain axis
compact_v1_1/addback_gemma26.json — each dropped L0 section added back to v1.1, chat axis
tasks_agent.py, tasks_premise.py, tasks_abstain.py — the scorers as used
PREDICTIONS_*.md — pre-registered predictions and outcomes
L0_compact_v1.json, compact_v1_1/L0_compact_v1_1.json — the prompt artifacts
README.md — launch flags per model

Each row file is a JSON list; rows carry arm/variant (system layer), probe/q, seed, score, detail, tokens, and the raw content (agent rows also carry tool_calls).

Related

  • Code and validation note: github.com/mobius-style/mmv (prompts/l0_essentials_compact_v1_1.json, docs/L0_ESSENTIALS_COMPACT_VALIDATION.md)
  • Base doctrine: L0 v8.4 / v8.4.1 (same repository)
Downloads last month
36
Total file size:
788 kB