hamishivi/rds-sels-tydiqa-shots-top326k

RDS+ Selected TydiQA shots 326k

This is the dataset (and associated scores) selected by RDS+ when selecting 326k samples using TydiQA few-shot samples. For more details, please see the paper Practical Large-Scale Data Selection for Instruction Tuning.

This dataset is selected from Tulu 2 unfiltered, and please see that page for more information on sources.

When finetuning a Llama 2 7b model on this data using the associated codebase and evaluating with the same codebase, the expected results are:

	MMLU	GSM8k	BBH	TydiQA	Codex	Squad	AlpacaEval	Average
MMLU shots 326k	51.69	34.04	45.30	56.07	32.43	88.75	33.09	48.77
GSM8K shots 326k	50.66	35.94	45.02	55.67	31.08	85.02	37.03	48.63
BBH shots 326k	49.36	34.34	44.44	51.99	28.38	81.44	35.51	46.49
TydiQA shots 326k (this data)	50.89	14.78	45.78	57.68	27.03	88.94	29.61	44.96
Codex 326k	47.17	35.25	45.10	52.09	33.78	83.35	45.01	48.82
Squad 326k	51.18	12.21	42.84	57.17	26.35	87.87	31.58	44.17
AlpacaFarm 326k	51.45	21.76	44.76	52.24	26.35	81.43	51.13	47.02

This is using the following hyperparameters:

• learning_rate: 2e-05
• total_train_batch_size: 128
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 2.0

License

We are releasing this dataset under the terms of ODC-BY. By using this, you are also bound by the Common Crawl terms of use in respect of the content contained in the dataset. Note that that different licenses apply to subsets of the data. See Tulu 2 unfiltered for details.

Citation

If you find this model or data is useful in your work, please cite it with:

@misc{ivison2025data,
      title={{Practical Large-Scale Data Selection for Instruction Tuning}}, 
      author={{Hamish Ivison and Muru Zhang and Faeze Brahman and Pang Wei Koh and Pradeep Dasigi}}
      year={2025},
      eprint={2503.01807},
      url={https://arxiv.org/abs/2503.01807},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}