The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
all_gold: string
raw_text: string
messages: list<element: struct<role: string, content: string>>
child 0, element: struct<role: string, content: string>
child 0, role: string
child 1, content: string
-- schema metadata --
huggingface: '{"info": {"features": {"all_gold": {"dtype": "string", "_ty' + 208
to
{'raw_text': Value('string'), 'messages': List({'content': Value('string'), 'role': 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/parquet/parquet.py", line 220, in _generate_tables
yield Key(file_idx, batch_idx), self._cast_table(pa_table)
~~~~~~~~~~~~~~~~^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/parquet/parquet.py", line 156, 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
all_gold: string
raw_text: string
messages: list<element: struct<role: string, content: string>>
child 0, element: struct<role: string, content: string>
child 0, role: string
child 1, content: string
-- schema metadata --
huggingface: '{"info": {"features": {"all_gold": {"dtype": "string", "_ty' + 208
to
{'raw_text': Value('string'), 'messages': List({'content': Value('string'), 'role': Value('string')})}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
BioNER Fixed Evaluation Subset
A fixed, pre-built evaluation subset for 9 standard biomedical Named
Entity Recognition (NER) benchmarks, created for reproducible pre/post
fine-tuning comparisons of MedGemma-4B (google/medgemma-1.5-4b-it) as
part of an internship project at DISI, University of Bologna.
Why this exists
Comparing a model before and after fine-tuning — and later against a further CoT-distilled version — is only meaningful if every evaluation uses exactly the same examples, in exactly the same prompt form. The underlying entity-distractor sampling used when converting IOB spans into prompts is itself randomized (see disi-unibo-nlp's BioNER Datasets collection and the SFT data pipeline this project builds on), so re-running the conversion script twice on the "same" rows can silently produce two different prompts. This dataset freezes that randomness once, so every future evaluation — today, and months from now — compares models on identical inputs.
What's in it
Nine configs, one per source dataset, each already converted into a ready-to-use chat-format prompt:
| Config | Source | Rows |
|---|---|---|
ncbi |
disi-unibo-nlp/ncbi | 500 |
bc5cdr |
disi-unibo-nlp/bc5cdr | 500 |
bc2gm |
disi-unibo-nlp/bc2gm | 500 |
bc4chemd |
disi-unibo-nlp/bc4chemd | 500 |
AnatEM |
disi-unibo-nlp/AnatEM | 500 |
JNLPBA |
disi-unibo-nlp/JNLPBA | 500 |
JNLPBA-Rare |
disi-unibo-nlp/JNLPBA-Rare | 465 (full test set — smaller than 500) |
MedMentions-Rare |
disi-unibo-nlp/MedMentions-Rare | 500 |
biored |
disi-unibo-nlp/biored | 500 |
Each row is a single split (test) with one column, messages: a
3-turn chat (system, user, assistant) in the exact format expected
by MedGemma's chat template. The assistant turn holds the gold-standard
answer (JSON object mapping entity types to verbatim text spans) used as
ground truth during evaluation — not meant to be shown to the model
at inference time (drop the last turn before generating).
How the subset was built
- 500 rows sampled per dataset (or the full test set, if smaller — see
JNLPBA-Rareabove), using a fixed random seed (42) over row indices. - For datasets without pre-existing validation/test splits, an 95/5 train/test split was applied (same seed), matching the split logic used during the model's own fine-tuning.
- Prompts were built with the same
SFTDataProcessor/IOBSpanExtractor/ entity-distractor-sampling logic used to prepare the SFT training data — reused verbatim, not reimplemented, to avoid any drift between training-time and evaluation-time prompt construction.
Licensing note
This is a derived subset of the datasets listed above, each carrying its own original license and citation requirements (see each source repo's own dataset card). No license is asserted over the source data itself here — check the original repos before reuse for anything beyond reproducing this project's own evaluation.
- Downloads last month
- 380