Datasets:

lamsheeper-data-attribution
/

ftrace-trex-results

ArXiv:

Tags:

data-attribution

influence-functions

interpretability

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

Duplicate

Split (1)

train

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
files: struct<cf50_facts.jsonl.gz: string, rankings/bm25.npz: string, rankings/layerwise_trak_top10_d9280.n (... 144 chars omitted)
  child 0, cf50_facts.jsonl.gz: string
  child 1, rankings/bm25.npz: string
  child 2, rankings/layerwise_trak_top10_d9280.npz: string
  child 3, rankings/layerwise_trak_top10_d9792.npz: string
  child 4, rankings/trak.npz: string
  child 5, rankings/trak_nlo.npz: string
  child 6, retrain_top1.npz: string
ranking_sources: struct<trak: struct<sha256: string, depth: int64>, trak_nlo: struct<sha256: string, depth: int64>, l (... 174 chars omitted)
  child 0, trak: struct<sha256: string, depth: int64>
      child 0, sha256: string
      child 1, depth: int64
  child 1, trak_nlo: struct<sha256: string, depth: int64>
      child 0, sha256: string
      child 1, depth: int64
  child 2, layerwise_trak_top10_d9792: struct<sha256: string, depth: int64>
      child 0, sha256: string
      child 1, depth: int64
  child 3, layerwise_trak_top10_d9280: struct<sha256: string, depth: int64>
      child 0, sha256: string
      child 1, depth: int64
  child 4, bm25: struct<sha256: string, depth: int64>
      child 0, sha256: string
      child 1, depth: int64
removal_sets: struct<2000: struct<bm25: string, gt: string, layerwise_trak_top10: string, trak: string, trak_nlo:  (... 213 chars omitted)
  child 0, 2000: struct<bm25: string, gt: string, layerwise_trak_top10: string, trak: string, trak_nlo: string>
      child 0, bm25: string
      child 1, gt: string
      child 2, layerwise_trak_top10: string
      child 3, trak: string
      child 4, trak_nlo: string
  child 1, 5000: struct<bm25: string, gt: string, layerwise_trak_top10: string, trak: string, trak_nlo: string>
      child 0, bm25: string
      child 1, gt: string
      child 2, layerwise_trak_top10: string
      child 3, trak: string
      child 4, trak_nlo: string
  child 2, 11522: struct<bm25: string, gt: string, layerwise_trak_top10: string, trak: string, trak_nlo: string>
      child 0, bm25: string
      child 1, gt: string
      child 2, layerwise_trak_top10: string
      child 3, trak: string
      child 4, trak_nlo: string
uuid: string
n_proponents: int64
n_proponent_sentences: int64
row: int64
ft_p_answer: double
predicate_id: string
fact_key: string
target_norm: string
proponent_rows: list<item: int64>
  child 0, item: int64
to
{'uuid': Value('string'), 'row': Value('int64'), 'predicate_id': Value('string'), 'fact_key': Value('string'), 'target_norm': Value('string'), 'n_proponents': Value('int64'), 'n_proponent_sentences': Value('int64'), 'ft_p_answer': Value('float64'), 'proponent_rows': List(Value('int64'))}
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 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
              files: struct<cf50_facts.jsonl.gz: string, rankings/bm25.npz: string, rankings/layerwise_trak_top10_d9280.n (... 144 chars omitted)
                child 0, cf50_facts.jsonl.gz: string
                child 1, rankings/bm25.npz: string
                child 2, rankings/layerwise_trak_top10_d9280.npz: string
                child 3, rankings/layerwise_trak_top10_d9792.npz: string
                child 4, rankings/trak.npz: string
                child 5, rankings/trak_nlo.npz: string
                child 6, retrain_top1.npz: string
              ranking_sources: struct<trak: struct<sha256: string, depth: int64>, trak_nlo: struct<sha256: string, depth: int64>, l (... 174 chars omitted)
                child 0, trak: struct<sha256: string, depth: int64>
                    child 0, sha256: string
                    child 1, depth: int64
                child 1, trak_nlo: struct<sha256: string, depth: int64>
                    child 0, sha256: string
                    child 1, depth: int64
                child 2, layerwise_trak_top10_d9792: struct<sha256: string, depth: int64>
                    child 0, sha256: string
                    child 1, depth: int64
                child 3, layerwise_trak_top10_d9280: struct<sha256: string, depth: int64>
                    child 0, sha256: string
                    child 1, depth: int64
                child 4, bm25: struct<sha256: string, depth: int64>
                    child 0, sha256: string
                    child 1, depth: int64
              removal_sets: struct<2000: struct<bm25: string, gt: string, layerwise_trak_top10: string, trak: string, trak_nlo:  (... 213 chars omitted)
                child 0, 2000: struct<bm25: string, gt: string, layerwise_trak_top10: string, trak: string, trak_nlo: string>
                    child 0, bm25: string
                    child 1, gt: string
                    child 2, layerwise_trak_top10: string
                    child 3, trak: string
                    child 4, trak_nlo: string
                child 1, 5000: struct<bm25: string, gt: string, layerwise_trak_top10: string, trak: string, trak_nlo: string>
                    child 0, bm25: string
                    child 1, gt: string
                    child 2, layerwise_trak_top10: string
                    child 3, trak: string
                    child 4, trak_nlo: string
                child 2, 11522: struct<bm25: string, gt: string, layerwise_trak_top10: string, trak: string, trak_nlo: string>
                    child 0, bm25: string
                    child 1, gt: string
                    child 2, layerwise_trak_top10: string
                    child 3, trak: string
                    child 4, trak_nlo: string
              uuid: string
              n_proponents: int64
              n_proponent_sentences: int64
              row: int64
              ft_p_answer: double
              predicate_id: string
              fact_key: string
              target_norm: string
              proponent_rows: list<item: int64>
                child 0, item: int64
              to
              {'uuid': Value('string'), 'row': Value('int64'), 'predicate_id': Value('string'), 'fact_key': Value('string'), 'target_norm': Value('string'), 'n_proponents': Value('int64'), 'n_proponent_sentences': Value('int64'), 'ft_p_answer': Value('float64'), 'proponent_rows': List(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.

FTRACE-TREX remove-and-retrain results

The records behind the FTRACE-TREX experiment of "Elucidating the Design Space of LM Data Attribution" (App. E: Table 16, Figure 7). Following the FTRACE-TREX setup of Akyürek et al. (2022) and the remove-and-retrain protocol of TRAK (App. H.4), mT5-small is fine-tuned on TREx abstract sentences. Each attribution method ranks the training corpus for 50 novel facts, the top-ranked documents are removed round-robin up to a budget B, and mT5-small is retrained without them at seeds 2001 and 2002. A larger drop in top-1 accuracy on the 50 facts means the method found the documents the model relied on.

Results

Drop in top-1 accuracy (percentage points) on the 50 facts, mean over the two retraining seeds ± one standard error. The no-removal control reaches 86% and 84% at seeds 2001 and 2002.

Removal ranking	B = 2,000	B = 5,000	B = 11,522
TRAK	5 ± 3	9 ± 5	9 ± 3
TRAK (negative log-odds)	15 ± 3	32 ± 4	47 ± 1
Layerwise TRAK, top 10 tokens	9 ± 1	19 ± 1	25 ± 3
BM25	3 ± 1	5 ± 1	19 ± 3
Ground-truth proponents	4 (one seed)	4 ± 4	6 ± 0

summaries/summary_b{B}.json holds these and every other number of Table 16: per-seed accuracies (acc), drops (drop, drop_mean, drop_se, a bootstrap drop_ci95), the drops on the other novel facts of the same relations and of other relations, and the mean fraction of each fact's proponents removed (gt_per_fact_removed_mean).

Files

summaries/summary_b{B}.json   the results above, one file per budget B
cf50_facts.jsonl.gz           the 50 facts: FTRACE uuid, predicate, normalised answer,
                              proponent training rows, fine-tuned model's p(answer)
retrain_top1.npz              top-1 correctness of every retrained model on every
                              novel fact: uuids [8437], predicate_ids [8437],
                              runs [35] ("{B}/{ranking}/{seed}"), top1 bool [35, 8437]
rankings/{ranking}.npz        each method's ranking of the training corpus for the
                              50 facts: query_uuids [50], rows int32 [50, depth]
manifest.json                 sha256 of every file, of the full rankings they were
                              cut from, and of the removal sets they rebuild

Training rows are positions in the abstracts configuration of ekinakyurek/ftrace (its parquet shards on revision refs/convert/parquet, concatenated in path order: 1,560,453 masked examples); uuids are those of its queries configuration. A ranking is cut to the depth the largest budget reaches; round-robin removal never reads further, so the cut rankings rebuild the removal sets exactly.

Notes

Ranking names. trak multiplies by TRAK's Q = 1 - p̄; trak_nlo is the same score without Q; layerwise_trak_top10 scores a document by its 10 largest per-position terms. The pipeline that produced them called them jl_w, jl_nores and trak_top10.
Top-10 feature widths. layerwise_trak_top10 uses D = 9,792 (linear modules plus the shared embedding) at B = 2,000 and 5,000, and D = 9,280 (linear modules only) at B = 11,522; both rankings are here.
One missing seed. The ground-truth removal model at B = 2,000, seed 2001 was trained but never evaluated, so that cell averages one seed.
35 results, 31 evaluation files. The no-removal controls are shared across budgets.

The code that produced these files (catt ftrace) is at tag v1.0-validated of the CATT release.

Downloads last month: 33

Total file size:

852 kB

Papers for lamsheeper-data-attribution/ftrace-trex-results

TRAK: Attributing Model Behavior at Scale

Paper • 2303.14186 • Published Mar 24, 2023

Towards Tracing Factual Knowledge in Language Models Back to the Training Data

Paper • 2205.11482 • Published May 23, 2022