Datasets:

Isaacdnigerian
/

plumb-runs

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
task_id: string
probe: string
status: string
parse_ok: bool
format_strict: bool
fabricated: bool
silent_guess: bool
collateral: bool
over_refused: bool
completed: bool
raw: string
n_tasks: int64
collateral_write_rate: double
parse_success_rate: double
fabrication_rate: double
over_refusal_rate: double
completion_rate: double
format_strict_rate: double
denominators: struct<unanswerable: int64, answerable: int64, ambiguous: int64, with_write_set: int64>
  child 0, unanswerable: int64
  child 1, answerable: int64
  child 2, ambiguous: int64
  child 3, with_write_set: int64
silent_guess_rate: double
to
{'n_tasks': Value('int64'), 'fabrication_rate': Value('float64'), 'silent_guess_rate': Value('float64'), 'collateral_write_rate': Value('float64'), 'over_refusal_rate': Value('float64'), 'completion_rate': Value('float64'), 'parse_success_rate': Value('float64'), 'format_strict_rate': Value('float64'), 'denominators': {'unanswerable': Value('int64'), 'answerable': Value('int64'), 'ambiguous': Value('int64'), 'with_write_set': Value('int64')}}
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
              task_id: string
              probe: string
              status: string
              parse_ok: bool
              format_strict: bool
              fabricated: bool
              silent_guess: bool
              collateral: bool
              over_refused: bool
              completed: bool
              raw: string
              n_tasks: int64
              collateral_write_rate: double
              parse_success_rate: double
              fabrication_rate: double
              over_refusal_rate: double
              completion_rate: double
              format_strict_rate: double
              denominators: struct<unanswerable: int64, answerable: int64, ambiguous: int64, with_write_set: int64>
                child 0, unanswerable: int64
                child 1, answerable: int64
                child 2, ambiguous: int64
                child 3, with_write_set: int64
              silent_guess_rate: double
              to
              {'n_tasks': Value('int64'), 'fabrication_rate': Value('float64'), 'silent_guess_rate': Value('float64'), 'collateral_write_rate': Value('float64'), 'over_refusal_rate': Value('float64'), 'completion_rate': Value('float64'), 'parse_success_rate': Value('float64'), 'format_strict_rate': Value('float64'), 'denominators': {'unanswerable': Value('int64'), 'answerable': Value('int64'), 'ambiguous': Value('int64'), 'with_write_set': Value('int64')}}
              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.

PLUMB: run artifacts, logs and traces

Every scored model reply behind every number in the PLUMB write-up. Lexsi AI Safety and Alignment Hackathon, Track 1, problem statement 3.

Task sets: https://huggingface.co/datasets/Isaacdnigerian/plumb-tasks
Code, scorer, write-up: https://github.com/Dprof-in-tech/plumb
Model and adapters: https://huggingface.co/Isaacdnigerian/plumb
Interactive explorer: https://huggingface.co/spaces/Isaacdnigerian/plumb-benchmark

This exists so the numbers can be checked rather than believed. Every table in the write-up is recomputable from these files with the scorer in the repo, and the write-up documents eleven invalid measurements caught during the project, so the raw replies matter more here than usual.

Layout

`runs/` (201 files)

One .jsonl per evaluation run, with a matching .metrics.json for most. Each line of a .jsonl is one task's raw model reply:

{"id": "curator-1-007", "raw": "{\"status\": \"answered\", ...}", "finish_reason": "stop"}

finish_reason is recorded because its absence caused one of the documented failures: an empty reply that was a truncated generation got read as a deliberate refusal, and the artifact had no field that could tell the two apart. Older files predating that fix do not carry it, and the scorer now prints a warning when it is missing rather than scoring them as clean.

Naming follows <experiment>-<condition>-<replicate>.jsonl:

prefix	what it is
`baseline-`, `ollama-q4-`	baseline scoring of the converted base model
`sweep-{global,fire,water,earth}-*`	the four Tiny Aya variants on the full benchmark, 2 replicates each
`tier-`, `tiersweep-`	the three-tier ambiguity probe (unannounced, announced, control)
`auto--unannounced-` / `auto--auto-announced-`	the headline before-and-after: the `plumb.ambiguity` detector off and on
`rec-`, `hyb-`, `bal-*`	training attempts 5, 6 and 7 against their converted-base controls
`dax1-probe-*`	the Dax-1 probe. `v3` is the reported run with the model card's documented decoding; `v2` is the same tasks with thinking on; the unsuffixed file is the superseded 36-task run
`auditkit-*`	AuditKIT cross-check, wired to the status field and to the text fallback

`curatorkit/` (9 directories)

CuratorKIT review and hygiene passes over the generated task candidates, kept because the benchmark's provenance is part of what is being claimed. Includes the review that failed to catch the seven mislabelled tasks, which is itself a documented finding.

Reproducing the headline result

git clone https://github.com/Dprof-in-tech/plumb && cd plumb
python3 scripts/paired_analysis.py \
  runs/auto-global-unannounced-2.jsonl \
  runs/auto-global-auto-announced-2.jsonl

Expected: silent_guess_rate 100.0% to 0.0%, 16 discordant pairs all improving, 0 regressing, exact McNemar p = 3.1e-05.

The Dax-1 probe:

python3 scripts/score_dax_probe.py \
  --tasks tasks/dax-probe-large.jsonl runs/dax1-probe-v3-results.jsonl

What is not here

runs/ollama-models/ from the source repo is excluded. It is an 8GB local Ollama blob cache, not a run artifact. Model weights live in the model repo.

Caveat

These are raw replies, not a curated dataset. Several runs are invalid and are kept deliberately, because the write-up's contribution is the eleven measurement failures and the control that caught each. A file being present here does not mean its numbers were ever valid. The write-up says which.