Elamur Research⌕
☰
home / papers / 2024 / The Unreasonable Effectiveness of Easy Training Data for Hard Tasks

The Unreasonable Effectiveness of Easy Training Data for Hard Tasks

Paper · 2024 · Peter Hase, Mohit Bansal, Peter Clark et al.

Summary

The paper addresses scalable oversight: how to supervise language models on hard questions when correct labels for such questions are expensive and error-prone. The authors measure easy-to-hard generalization, meaning how well models fitted only on easy examples answer hard ones. They compare this against a zero-shot lower bound and a hard-trained oracle using a metric called Supervision Gap Recovered (SGR). Experiments use ARC, MMLU, GSM8k and StrategyQA, open models from 7b to 70b (Llama-2, Mixtral-8x7B), and three fitting methods: in-context learning (with and without chain-of-thought), QLoRA, and linear classifier heads. Difficulty is defined by seven hardness measures, six human-based and one loss-based. Across these settings, easy supervision typically recovers 70–100% of the gap, so it is nearly as useful as hard supervision. When hard labels cost more or contain more errors, training on easy data can even yield better hard-test accuracy. The authors conclude that scalable oversight may be less daunting than assumed, although generalization weakens when the train–test difficulty gap becomes very large.

Key points

  • SGR is defined as (Easy − Unsupervised) / (Hard − Unsupervised). It is usually 70–100% and stays near 100% for Llama-2 at 7b, 13b and 70b.
  • The six human hardness measures are an item's school grade, how difficult experts judge it to be, the cognitive abilities it demands, the length of the question, the length of the answer, and how many reasoning steps it involves. These measures disagree about which items are hardest.
  • In a cost scenario where hard labels are twice as expensive, which matches MMLU-STEM-5 (603 college vs. 1143 high-school questions), easy data more often than not gives better hard-test accuracy. Easy data also wins when its label error is 10% or more against 20% or more for hard data.
  • On college STEM questions with ICL, SGR falls from 74% when prompting with 8th-grade ARC questions to 57% with 3rd-grade questions. Very easy supervision is therefore weaker, but it still helps.
  • Code is released at github.com/allenai/easy-to-hard-generalization.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🔬Allen Institute for Artificial IntelligenceSeattle · United States317 works · 106,633 citations🎓University of North Carolina at Chapel HillChapel Hill · United States77 works · 23,365 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 19,503 worksArtificial Intelligence · 18,805 worksArts and Humanities · 17,959 worksLanguage and Linguistics · 17,789 works

Cite

Cite as Peter Hase, Mohit Bansal, Peter Clark and Sarah Wiegreffe. 2024. The Unreasonable Effectiveness of Easy Training Data for Hard Tasks. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2024.acl-long.378 · arXiv 2401.06751

@inproceedings{hase2024unreasonable,
  title = {{The Unreasonable Effectiveness of Easy Training Data for Hard Tasks}},
  author = {Peter Hase, Mohit Bansal, Peter Clark and Sarah Wiegreffe},
  booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2024},
  doi = {10.18653/v1/2024.acl-long.378},
  eprint = {2401.06751},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/2024.acl-long.378}
}

References

48 references, 25 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Nancy E Adams. 2015. Bloom’s taxonomy of cognitive learning objectives. Journal of the Medical Library Association: JMLA, 103(3):152. find
  2. Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. 2016. Concrete problems in ai safety. arXiv preprint arXiv:1606.06565.
  3. Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, P… — Qwen Technical Report
  4. Arpit Bansal, Avi Schwarzschild, Eitan Borgnia, Zeyad Emam, Furong Huang, Micah Goldblum, and Tom Goldstein. 2022. End-to-end algorithm synthesis with recurrent networks: Extrapolation without overthinking. Advances in Neural Information Processing Systems, 35:20232–20242. find
  5. Yonatan Belinkov. 2022. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207–219.
  6. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48.
  7. Benjamin S Bloom, Max D Engelhart, Edward J Furst, Walker H Hill, and David R Krathwohl. 1956. Taxonomy of educational objectives: The classification of educational goals. Handbook 1: Cognitive domain. McKay New York. find
  8. Ben Bogin, Shivanshu Gupta, and Jonathan Berant. 2022. Unobserved local structures make compositional generalization hard. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2731–2747, Abu Dhabi, United Arab Emirates. Association for Computational Lingui… find
  9. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901.
  10. Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, et al. 2023. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. arXiv preprint arXiv:2312.09390.
  11. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. ArXiv, abs/1803.05457.
  12. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
  13. Tim Dettmers, Mike Lewis, Sam Shleifer, and Luke Zettlemoyer. 2022. 8-bit optimizers via block-wise quantization. 9th International Conference on Learning Representations, ICLR.
  14. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLoRA: Efficient finetuning of quantized LLMs. In Thirty-seventh Conference on Neural Information Processing Systems.
  15. Anca Dumitrache, Lora Aroyo, and Chris Welty. 2018. Crowdsourcing ground truth for medical relation extraction. ACM Transactions on Interactive Intelligent Systems (TiiS), 8(2):1–20. find
  16. Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D Hwang, et al. 2023. Faith and fate: Limits of transformers on compositionality. In Thirty-seventh Conference on Neural Information Processing Systems.
  17. Bradley Efron and Robert J Tibshirani. 1994. An Introduction to the Bootstrap. CRC press. find
  18. Erich Elsen, Augustus Odena, Maxwell Nye, Sağ-nak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, and Arushi Somani. 2023. Releasing Persimmon-8B. Blogpost, ADEPT AI. find
  19. Yao Fu, Hao Peng, Ashish Sabharwal, Peter Clark, and Tushar Khot. 2023. Complexity-based prompting for multi-step reasoning. In The Eleventh International Conference on Learning Representations.
  20. Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346– 361. find
  21. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. In International Conference on Learning Representations.
  22. Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock,… find
  23. Brenden Lake and Marco Baroni. 2018. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International Conference on Machine Learning, pages 2873–2882. PMLR. find
  24. Matthew Lease. 2011. On quality control and machine learning in crowdsourcing. In Workshops at the twenty-fifth AAAI conference on artificial intelligence. Citeseer. find
  25. Eric Lehman, Jay DeYoung, Regina Barzilay, and Byron C. Wallace. 2019. Inferring which medical treatments work from reports of clinical trials. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1…
  26. Kevin Liu, Stephen Casper, Dylan Hadfield-Menell, and Jacob Andreas. 2023. Cognitive dissonance: Why do language model outputs disagree with internal representations of truthfulness? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4791–4797, Singapore… find
  27. Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. 2022. Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft. ↗
  28. Samuel Marks and Max Tegmark. 2023. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824.
  29. Mistral AI. 2023. Announcing Mistral 7B. Blogpost. find
  30. Yixin Nie, Xiang Zhou, and Mohit Bansal. 2020. What can we learn from collective human opinions on natural language inference data? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9131–9143, Online. Association for Computational Linguistics.
  31. Curtis G Northcutt, Anish Athalye, and Jonas Mueller. 2021. Pervasive label errors in test sets destabilize machine learning benchmarks. arXiv preprint arXiv:2103.14749. find
  32. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soum… — PyTorch: An Imperative Style, High-Performance Deep Learning Library
  33. Ellie Pavlick and Tom Kwiatkowski. 2019. Inherent disagreements in human textual inferences. Transactions of the Association for Computational Linguistics, 7:677–694.
  34. Ethan Perez, Douwe Kiela, and Kyunghyun Cho. 2021. Rissanen data analysis: Examining dataset characteristics via description length. In International Conference on Machine Learning, pages 8500–8513. PMLR. find
  35. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. 2023. Gpqa: A graduate-level google-proof q&a benchmark. arXiv preprint arXiv:2311.12022.
  36. Swarnadeep Saha, Peter Hase, and Mohit Bansal. 2023. Can language models teach weaker agents? teacher explanations improve students via theory of mind. In Thirty-seventh Conference on Neural Information Processing Systems. find
  37. Nithya Sambasivan, Shivani Kapania, Hannah Highfill, Diana Akrong, Praveen Paritosh, and Lora M Aroyo. 2021. “everyone wants to do the model work, not the data work”: Data cascades in high-stakes ai. In proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1–15. find
  38. Avi Schwarzschild, Eitan Borgnia, Arjun Gupta, Furong Huang, Uzi Vishkin, Micah Goldblum, and Tom Goldstein. 2021. Can you learn an algorithm? generalizing from easy to hard problems with recurrent networks. Advances in Neural Information Processing Systems, 34:6695–6706. find
  39. Charles Spearman. 1987. The proof and measurement of association between two things. The American journal of psychology, 100(3/4):441–471. find
  40. Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pa… find
  41. TII. 2023. Falcon LLM. Blogpost. find
  42. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023a. Llama: Open and efficient foundation language models.
  43. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023b. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.
  44. Elena Voita and Ivan Titov. 2020. Information-theoretic probing with minimum description length. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 183–196, Online. Association for Computational Linguistics.
  45. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, pages 24824–24837. Curran Associates, Inc.
  46. Benfeng Xu, Licheng Zhang, Zhendong Mao, Quan Wang, Hongtao Xie, and Yongdong Zhang. 2020. Curriculum learning for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6095–6104.
  47. Hattie Zhou, Arwen Bradley, Etai Littwin, Noam Razin, Omid Saremi, Josh Susskind, Samy Bengio, and Preetum Nakkiran. 2023a. What algorithms can transformers learn? a study in length generalization. arXiv preprint arXiv:2310.16028.
  48. Xiang Zhou, Yichen Jiang, and Mohit Bansal. 2023b. Data factors for better compositional generalization. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14549–14566, Singapore. Association for Computational Linguistics. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/hase-2024-unreasonable-effectiveness-easy-training-data/