Datasets:
The dataset viewer is not available for this split.
Error code: FeaturesError
Exception: ArrowInvalid
Message: Schema at index 1 was different:
model: string
lang: string
n: int64
coverage: double
selective_acc: double
hallucination_rate: double
native_refusal_share: double
vs
model: string
lang: string
n: int64
coverage: double
selective_acc: double
hallucination_rate: double
native_refusal_share: double
over_refusal: double
abstain_recall: double
abstain_precision: double
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 244, in compute_first_rows_from_streaming_response
iterable_dataset = iterable_dataset._resolve_features()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 4645, in _resolve_features
features = _infer_features_from_batch(self.with_format(None)._head())
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2866, in _head
return next(iter(self.iter(batch_size=n)))
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3073, in iter
for key, pa_table in ex_iterable.iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2581, in _iter_arrow
yield from self.ex_iterable._iter_arrow()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 606, in _iter_arrow
yield new_key, pa.Table.from_batches(chunks_buffer)
~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 5012, in pyarrow.lib.Table.from_batches
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: Schema at index 1 was different:
model: string
lang: string
n: int64
coverage: double
selective_acc: double
hallucination_rate: double
native_refusal_share: double
vs
model: string
lang: string
n: int64
coverage: double
selective_acc: double
hallucination_rate: double
native_refusal_share: double
over_refusal: double
abstain_recall: double
abstain_precision: doubleNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
TinyAya IDK: data, labels, outputs
Data for the Lexsi AI Safety & Alignment Hackathon (Track 1) project "Teaching TinyAya to say 'I don't know' in languages with no labelled QA data". Code: https://github.com/Eman2123/tinyaya-idk | Adapters: https://huggingface.co/EmanMirza2025/tinyaya-idk-adapters
Contents
data/raw/{en,es,ar,ms}.jsonl: 1500 MKQA factoid questions per language (same ids across languages; single-answer questions only).trap_{lang}.jsonl: teacher-written trap questions for recipe R1 (written by Claude Haiku via AI/ML API; a non-Cohere model used only as teacher).outputs/labels/{lang}.jsonl: per question: greedy answer, 5 sampled answers, agreement count, gold-known label, consistency label, probe score. Splits: build 995 / val 253 / test 252.data/processed/R1..R5/{lang}.jsonl: SFT training sets. R1 teacher traps, R2 gold labels (upper bound), R3 consistency labels, R4 transferred-probe labels, R5 probe AND consistency.outputs/results/: per-run evaluation CSVs (eval_R{k}_{lang}_s{seed}.csv), AuditKIT output, CuratorKIT report.data/curated_out/: CuratorKIT manifests, provenance, checksums and dataset cards per recipe/language.
Method in one paragraph
known/unknown labels come from (a) consistency of 5 samples (>=4 agree) and (b) a logistic-regression probe on the layer -12 hidden state of the last prompt token, trained on English gold labels only and applied to es/ar/ms. Model: CohereLabs/tiny-aya-global.
Licence and attribution
Questions and gold answers derive from MKQA (Apple, CC BY-SA 3.0), so this dataset is released under CC BY-SA 3.0. Model outputs were generated with CohereLabs/tiny-aya-global; check that model's licence for output use.
Limitations
Single QA source (MKQA, parallel questions); gold matching is normalized exact match with year-level date matching, so some correct answers count as wrong and MKQA gold labels are sometimes themselves wrong; Arabic/Malay golds have fewer aliases than English; refusal templates (es No lo sé., ar لا أعرف., ms Saya tidak tahu.) and few-shot prompts were not checked by native speakers; abstention tuning causes over-refusal (19-38% of answerable questions refused).
- Downloads last month
- 94