Datasets:
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: CastError
Message: Couldn't cast
pairs: struct<A: list<item: list<item: double>>, B: list<item: list<item: double>>, C: list<item: list<item (... 11 chars omitted)
child 0, A: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
child 1, B: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
child 2, C: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
label_mass: list<item: double>
child 0, item: double
A_understanding_probe: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
B_tone_probe: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
C_factuality_probe: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
to
{'A_understanding_probe': List(List(Value('string'))), 'B_tone_probe': List(List(Value('string'))), 'C_factuality_probe': List(List(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 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
pairs: struct<A: list<item: list<item: double>>, B: list<item: list<item: double>>, C: list<item: list<item (... 11 chars omitted)
child 0, A: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
child 1, B: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
child 2, C: list<item: list<item: double>>
child 0, item: list<item: double>
child 0, item: double
label_mass: list<item: double>
child 0, item: double
A_understanding_probe: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
B_tone_probe: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
C_factuality_probe: list<item: list<item: string>>
child 0, item: list<item: string>
child 0, item: string
to
{'A_understanding_probe': List(List(Value('string'))), 'B_tone_probe': List(List(Value('string'))), 'C_factuality_probe': List(List(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.
Non-Native (L2) English Capability Gap Evaluation
This repository contains an evaluation suite designed to assess how Qwen3.5-4B handles common grammar, syntax, and transfer patterns typically produced by L2 (non-native) English speakers.
1. The Underexplored Blind Spot: L2 English Pragmatic Bias
Standard benchmarks universally evaluate models on fluent, standard English (such as academic prose, clean code, or curated dialogues). However, hundreds of millions of global internet users speak English as a Second Language (L2). This demographic regularly uses structures influenced by their native languages (e.g., direct translation of verbs, missing definite/indefinite articles, or atypical sentence word orders).
The Capabillity Gap
Our evaluation highlights a structural blind spot in contemporary LLMs:
- Pragmatic Misinterpretation: Direct L2 phrasing (e.g., "Send me the report, please" vs. "Could you please send me the report?") is disproportionately flagged as rude, blunt, or aggressive by models, even when the underlying intent is entirely polite.
- Factuality & Reasoning Penalty: Models show degraded accuracy or yield lower evaluation grades when identical prompts are framed with minor non-native syntactic errors, punishing non-native speakers for their expression rather than factual understanding.
2. Evaluation Results Summary (Qwen3.5-4B)
- Probe A (Understanding Accuracy): Native phrasing got 88% correctness, whereas L2 phrasing got 89%.
- Probe B (Tone Friendliness - 1 - P(rude)): Native phrasing score was 0.98, dropping sharply to 0.83 for L2 phrasing.
- Probe C (Factuality Grade 1-9): Native explanations received an average grade of 8.64, falling to 7.83 under L2 syntax constraints.
3. Path Forward: Closing the L2 Gap
To bridge this capability gap, we propose three fundamental adjustments:
- Pragmatic-Aware Curation (Pre-training / SFT): Intentionally curate training corpora to include diverse English dialects, international forums, and L2-rich text. Training models to recognize that syntactic variation does not imply impoliteness or logical failure.
- Contrastive Syntactic Fine-Tuning: Apply Contrastive Language-Image Pre-training (CLIP)-style or DPO objectives using pairs of equivalent Native/L2 phrasings to enforce representation invariance across grammatical variations.
- De-biased Reinforcement Learning: Calibrate RLHF reward models so they do not associate non-native phrasing or grammatical structure with lower safety, lower politeness, or lower quality scores.
- Downloads last month
- 3