Hugging Face's logo

Datasets:
M1ztyk
/
saige-2x2-ablation-results

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
config: struct<base_model: string, data: string, eval_records: int64, grouped_split_seed: int64, adapters: s (... 75 chars omitted)
  child 0, base_model: string
  child 1, data: string
  child 2, eval_records: int64
  child 3, grouped_split_seed: int64
  child 4, adapters: struct<run3: string, run4: string, run5: string>
      child 0, run3: string
      child 1, run4: string
      child 2, run5: string
  child 5, margin_definition: string
cells: struct<base|rs: struct<mean_margin: double, accuracy: double, per_record: list<item: double>>, base| (... 1197 chars omitted)
  child 0, base|rs: struct<mean_margin: double, accuracy: double, per_record: list<item: double>>
      child 0, mean_margin: double
      child 1, accuracy: double
      child 2, per_record: list<item: double>
          child 0, item: double
  child 1, base|generic: struct<mean_margin: double, accuracy: double, per_record: list<item: double>>
      child 0, mean_margin: double
      child 1, accuracy: double
      child 2, per_record: list<item: double>
          child 0, item: double
  child 2, base|none: struct<mean_margin: double, accuracy: double, per_record: list<item: double>>
      child 0, mean_margin: double
      child 1, accuracy: double
      child 2, per_record: list<item: double>
          child 0, item: double
  child 3, run3|rs: struct<mean_margin: double, accuracy: double, per_record: list<item: double>, delta_vs_base: double>
      child 0, mean_margin: double
      child 1, accuracy: doubl
...
d 3, acc: double
  child 1, run3: struct<rs: struct<mean: double, std: double, n: int64, acc: double>, generic: struct<mean: double, s (... 99 chars omitted)
      child 0, rs: struct<mean: double, std: double, n: int64, acc: double>
          child 0, mean: double
          child 1, std: double
          child 2, n: int64
          child 3, acc: double
      child 1, generic: struct<mean: double, std: double, n: int64, acc: double>
          child 0, mean: double
          child 1, std: double
          child 2, n: int64
          child 3, acc: double
      child 2, none: struct<mean: double, std: double, n: int64, acc: double>
          child 0, mean: double
          child 1, std: double
          child 2, n: int64
          child 3, acc: double
  child 2, run4: struct<rs: struct<mean: double, std: double, n: int64, acc: double>, generic: struct<mean: double, s (... 99 chars omitted)
      child 0, rs: struct<mean: double, std: double, n: int64, acc: double>
          child 0, mean: double
          child 1, std: double
          child 2, n: int64
          child 3, acc: double
      child 1, generic: struct<mean: double, std: double, n: int64, acc: double>
          child 0, mean: double
          child 1, std: double
          child 2, n: int64
          child 3, acc: double
      child 2, none: struct<mean: double, std: double, n: int64, acc: double>
          child 0, mean: double
          child 1, std: double
          child 2, n: int64
          child 3, acc: double
to
{'meta': {'base': Value('string'), 'adapters': {'run3': Value('string'), 'run4': Value('string')}, 'n_records': Value('int64'), 'n_eval_rows': Value('int64'), 'conditions': List(Value('string'))}, 'cells': {'base': {'rs': List(Value('float64')), 'generic': List(Value('float64')), 'none': List(Value('float64'))}, 'run3': {'rs': List(Value('float64')), 'generic': List(Value('float64')), 'none': List(Value('float64'))}, 'run4': {'rs': List(Value('float64')), 'generic': List(Value('float64')), 'none': List(Value('float64'))}}, 'records': List({'state': Value('string'), 'id': Value('int64'), 'native': List(Value('string')), 'rs': Value('float64'), 'generic': Value('float64'), 'none': Value('float64')}), 'summary': {'base': {'rs': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'generic': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'none': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}}, 'run3': {'rs': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'generic': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'none': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}}, 'run4': {'rs': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'generic': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'none': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}}}}
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
              config: struct<base_model: string, data: string, eval_records: int64, grouped_split_seed: int64, adapters: s (... 75 chars omitted)
                child 0, base_model: string
                child 1, data: string
                child 2, eval_records: int64
                child 3, grouped_split_seed: int64
                child 4, adapters: struct<run3: string, run4: string, run5: string>
                    child 0, run3: string
                    child 1, run4: string
                    child 2, run5: string
                child 5, margin_definition: string
              cells: struct<base|rs: struct<mean_margin: double, accuracy: double, per_record: list<item: double>>, base| (... 1197 chars omitted)
                child 0, base|rs: struct<mean_margin: double, accuracy: double, per_record: list<item: double>>
                    child 0, mean_margin: double
                    child 1, accuracy: double
                    child 2, per_record: list<item: double>
                        child 0, item: double
                child 1, base|generic: struct<mean_margin: double, accuracy: double, per_record: list<item: double>>
                    child 0, mean_margin: double
                    child 1, accuracy: double
                    child 2, per_record: list<item: double>
                        child 0, item: double
                child 2, base|none: struct<mean_margin: double, accuracy: double, per_record: list<item: double>>
                    child 0, mean_margin: double
                    child 1, accuracy: double
                    child 2, per_record: list<item: double>
                        child 0, item: double
                child 3, run3|rs: struct<mean_margin: double, accuracy: double, per_record: list<item: double>, delta_vs_base: double>
                    child 0, mean_margin: double
                    child 1, accuracy: doubl
              ...
              d 3, acc: double
                child 1, run3: struct<rs: struct<mean: double, std: double, n: int64, acc: double>, generic: struct<mean: double, s (... 99 chars omitted)
                    child 0, rs: struct<mean: double, std: double, n: int64, acc: double>
                        child 0, mean: double
                        child 1, std: double
                        child 2, n: int64
                        child 3, acc: double
                    child 1, generic: struct<mean: double, std: double, n: int64, acc: double>
                        child 0, mean: double
                        child 1, std: double
                        child 2, n: int64
                        child 3, acc: double
                    child 2, none: struct<mean: double, std: double, n: int64, acc: double>
                        child 0, mean: double
                        child 1, std: double
                        child 2, n: int64
                        child 3, acc: double
                child 2, run4: struct<rs: struct<mean: double, std: double, n: int64, acc: double>, generic: struct<mean: double, s (... 99 chars omitted)
                    child 0, rs: struct<mean: double, std: double, n: int64, acc: double>
                        child 0, mean: double
                        child 1, std: double
                        child 2, n: int64
                        child 3, acc: double
                    child 1, generic: struct<mean: double, std: double, n: int64, acc: double>
                        child 0, mean: double
                        child 1, std: double
                        child 2, n: int64
                        child 3, acc: double
                    child 2, none: struct<mean: double, std: double, n: int64, acc: double>
                        child 0, mean: double
                        child 1, std: double
                        child 2, n: int64
                        child 3, acc: double
              to
              {'meta': {'base': Value('string'), 'adapters': {'run3': Value('string'), 'run4': Value('string')}, 'n_records': Value('int64'), 'n_eval_rows': Value('int64'), 'conditions': List(Value('string'))}, 'cells': {'base': {'rs': List(Value('float64')), 'generic': List(Value('float64')), 'none': List(Value('float64'))}, 'run3': {'rs': List(Value('float64')), 'generic': List(Value('float64')), 'none': List(Value('float64'))}, 'run4': {'rs': List(Value('float64')), 'generic': List(Value('float64')), 'none': List(Value('float64'))}}, 'records': List({'state': Value('string'), 'id': Value('int64'), 'native': List(Value('string')), 'rs': Value('float64'), 'generic': Value('float64'), 'none': Value('float64')}), 'summary': {'base': {'rs': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'generic': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'none': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}}, 'run3': {'rs': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'generic': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'none': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}}, 'run4': {'rs': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'generic': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}, 'none': {'mean': Value('float64'), 'std': Value('float64'), 'n': Value('int64'), 'acc': Value('float64')}}}}
              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.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

SAIGE 2×2 ablation results

Logprob-margin evaluations of the SAIGE DPO adapters: for each state (base model + adapters) × prompt condition (rs / generic / none) × eval record, the margin is

margin = logp(chosen | prompt) − logp(rejected | prompt)

with per-token-normalized margins also reported (rules out length artifacts). Eval records are the 11 unique records of run4's grouped split (seed 42) of M1ztyk/SAIGE-right-speech-dpo — 17 rows grouped by chosen text so no chosen/rejected twin straddles the split. Missing prompt conditions are synthesized from the canonical RS/generic system prompts. Preference accuracy is sign(margin) per record.

Files

File What it is
ablation_v2.py Canonical scoring script — base/run3/run4/run5 in one job (internally consistent), raw + per-token margins, ABLATION_DRY_RUN=1 validates data prep
ablation_results_v2.json Canonical results — all four states scored fresh in a single run
ablation_run5.py Earlier-session script; differs only in adapter id (SAIGE-dpo-v4-run5) and output filename. Superseded
ablation_results_run5.json Output of ablation_run5.py — numbers agree with v2 within noise (independent reproducibility check)
ablation_results.json v1 ablation (base/run3/run4 only, before run5 existed; earlier scoring implementation — do not compare numbers across v1 and v2 tables, only within one)

Canonical results (ablation_results_v2.json)

Mean margin (nats) over 11 held-out records; per-token margins tell the same story:

State RS prompt Generic None
base −12.09 −15.62 −14.61
run3 −5.99 −10.41 −9.43
run4 −10.45 −14.05 −13.29
run5 −12.21 −15.56 −15.30

Findings

  1. All cells negative: no adapter flips preference on held-out records; the base model's prior against the chosen style is 12–16 nats deep.
  2. run3 moved most (+6.1 nats under RS) but did so with the weaker data; run4 moved less with the same data; run5 (16 high-contrast pairs) moved not at all on held-out records despite train margins of 2.49 — memorization, not generalization.
  3. The rs/generic gap stays flat for every adapter (base 6.1 → run4 6.3 → run3 7.3; run5 unchanged): adapters shift all conditions in parallel. The prompt-independent "experiential" mechanism is real; what's missing is data volume. Next: 9×9 expansion → retrain → re-ablate.
Downloads last month
59
Total file size:
42.7 kB