Jotschi/coco-karpathy-simple-en

Dataset Card for MS COCO Karpathy in Simple English

This dataset contains captions that were rephrased into simple english so that a young child would understand it.

Dataset Details

Dataset Sources

The processed MS COCO datasets (Karpathy Split) in this repo are based on the following sources:

Type	MD5	URL
Train	aa31ac474cf6250ebb81d18348a07ed8	https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_train.json
Validation	b273847456ef5580e33713b1f7de52a0	https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val.json
Test	3ff34b0ef2db02d01c37399f6a2a6cd1	https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test.json

MS COCO:

• Download: https://cocodataset.org/#download
• Paper: http://arxiv.org/abs/1405.0312

Dataset Creation

This dataset was generated by processing the annotations via Mistal7B.

Prompt used:

Rewrite the sentence " + caption + " for a 3 to 4 year old child. Give only one simple sentence. Don't use the word see. Give only a single answer.

A filter was applied to only store captions which matched the common output format. A best effort filter was applied to reduce the chance of including multiple example sentences in the output.

Curation Rationale

This dataset is useful for experiments with small LLMs which have only a reduced corpus. The dataset is suitable to be used for LAVIS experiments (QFormer Training) with a finetuned TinyStories 33M LLM.