veyra-ai/SciCloze-900
SciCloze-900 SciCloze-900 is a cloze-style GCSE Combined Science benchmark for evaluating small base language models. The benchmark contains 900 multiple-choice cloze items: 300 biology items 300 chemistry items 300 physics items Each item is designed as a natural text continuation rather than an instruction-style question. The intended evaluation method is to score each answer choice by average log probability per token as a continuation of the prompt. Splits… See the full description on the dataset page: https://huggingface.co/datasets/veyra-ai/SciCloze-900.
SciCloze-900
SciCloze-900 is a cloze-style GCSE Combined Science benchmark for evaluating small base language models.
The benchmark contains 900 multiple-choice cloze items:
- 300 biology items
- 300 chemistry items
- 300 physics items
Each item is designed as a natural text continuation rather than an instruction-style question. The intended evaluation method is to score each answer choice by average log probability per token as a continuation of the prompt.
Splits
validation: 900 items
The subject subsets can be recovered with the subject field:
BioCloze-300: rows wheresubject == "biology"ChemCloze-300: rows wheresubject == "chemistry"PhysCloze-300: rows wheresubject == "physics"
Notes
The benchmark items were AI-assisted and then validated. Source curriculum text was used to define topic scope; the released items are original cloze-style benchmark items and do not reproduce source text except for unavoidable scientific terminology
SciCloze-900 is released under the OGL v3.0 License.
To preserve benchmark comparability, modified versions of the dataset should be clearly renamed and should not be reported as official SciCloze-900 results.