STAIR (STructure Aware Information Retriever): A novel dataset and LLM based retriever for document structure augmentation
Paper • 2609.03874 • Published • 1
Dataset Viewer
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: TypeError
Message: Couldn't cast array of type string to null
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 2312, in cast_table_to_schema
cast_array_to_feature(
~~~~~~~~~~~~~~~~~~~~~^
table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
feature,
^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2068, in cast_array_to_feature
_c(array.field(name) if name in array_fields else null_array, subfeature)
~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2118, in cast_array_to_feature
casted_array_values = _c(array.values, feature.feature)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2152, in cast_array_to_feature
return array_cast(
array,
...<2 lines>...
allow_decimal_to_str=allow_decimal_to_str,
)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
return func(array, *args, **kwargs)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2014, in array_cast
raise TypeError(f"Couldn't cast array of type {_short_str(array.type)} to {_short_str(pa_type)}")
TypeError: Couldn't cast array of type string to nullNeed 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.
searchtome-subset — rebuilt STAIR/SearchTome benchmark (scaled-down)
This repository is an independent, scaled-down rebuild of the SearchTome
benchmark from STAIR: STructure Aware Information Retriever (arXiv
2609.03874, IBM, 2026). The paper's original
benchmark and code (linked as anonymous.4open.science/r/s_331) are no longer
available (expired anonymous repository); nothing was on the Hub. This rebuild
follows the paper's described construction pipeline.
Contents
data/<domain>/<book_id>/toc.json— ToC extracted with PyMuPDF from the open textbook PDFs (Open Textbook Library sources named in the paper's Table 1).data/<domain>/<book_id>/content.jsonl— parsed text blocks{id, page, section_path, text}.gen/<domain>/<book_id>/leaf_docs.jsonl— leaf sections (retrieval corpus):{leaf_id, title, page_start, page_end, text}.gen/<domain>/<book_id>/questions.jsonl— teacher-generated queries:{book_id, split, query, leaf_id, leaf_title, paragraph_id}.gen/raw_responses.jsonl— all raw teacher responses (8,890 prompts).eval/preds_stair.json,eval/preds_dsi.json— constrained-decoding predictions of the two finetuned models.
Books (2 domains × 3): Education — Open Music Theory (489 leaves), The Whole Child (140), Teaching in a Digital Age (172); Social Sciences — Aural Skills (131), Comparative Government (336), Behavioral Economics (208).
Questions: 16,993 total (2 per eligible paragraph, ≥200 chars after block merging), generated by mistralai/Mixtral-8x7B-Instruct-v0.1 (paper-exact, bf16, 4×A100 via vLLM). Split ≈55/33/12 test/train/dev per the paper's Table 1 convention, assigned per prompt with a fixed seed. Train = 5,344 questions.
Models (independent reproduction)
- STAIR LoRA:
SGK86/stair-mistral-7b-searchtome-subset(Mistral-7B-Instruct-v0.2, LoRA r=16 α=32, ToC in prompt, max_length 8192, 1 epoch / 667 steps, eval_loss 0.208) - DSI full-FT:
SGK86/dsi-mistral-7b-searchtome-subset(same data without ToC block, max_length 512, 1 epoch / 334 steps)
Results (100 test questions per book, seed 42, n=600; greedy constrained decoding, top-1 candidate → R@3 = R@1)
| Method | R@1 | R@3 | nDCG@3 |
|---|---|---|---|
| STAIR (ours) | 0.325 | 0.325 | 0.325 |
| DSI (ours) | 0.165 | 0.165 | 0.165 |
| BM25 (same sample) | 0.640 | 0.788 | 0.728 |
| STAIR (paper, 18 books) | 0.826 | 0.908 | 0.875 |
| DSI (paper) | 0.769 | 0.853 | 0.819 |
| BM25 (paper) | ~0.595 | — | — |
Directionally reproduced: STAIR > DSI (0.325 vs 0.165) — ToC conditioning helps — and STAIR produced zero empty/unparseable outputs vs 6.3 % for DSI (paper reports hallucination <0.05 %). NOT reproduced: absolute levels are far below the paper (1 epoch, 6 books, 5.3k train questions vs. paper's full corpus and up-to-200-epoch training), and BM25 outperforms both finetuned models in our rebuild — our Mixtral-generated queries are lexically much closer to their source sections than the paper's queries, inverting the paper's ranking. Detailed divergence list in the reproduction report.
Known divergences from the paper
- Original artifacts unavailable (expired anonymous repo) → full rebuild.
- 6 books / 2 domains instead of 18 / 6; 1 training epoch instead of up to 200 with early stopping.
- 2 questions/paragraph (paper: variable 0.6–3.9) → easier for BM25.
- Leaf counts from our ToC cleaning are higher than the paper's Table 1.
- Eval produces a single constrained candidate (greedy); the paper's R@3 > R@1 implies multiple candidates (beam), which we did not run.
- Zero-shot Mistral baseline not run.
Scripts: SGK86/stair-repro-code (dataset repo). Training dashboard:
SGK86/stair-repro-searchtome-trackio.
- Downloads last month
- 140