{"id":"arxiv:2406.11813","title":"How Do Large Language Models Acquire Factual Knowledge During Pretraining?","date":"2024-06-17","section":"Reasoning and the \"physics\" of language models","quality":0.09754,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/","source_url":"https://arxiv.org/abs/2406.11813","text_available":true,"authors":["Hoyeon Chang","Jinho Park","Seonghyeon Ye","Sohee Yang"],"outgoing":[{"source":"arxiv:2406.11813","target":"arxiv:2001.08361","evidence":"title","id":"arxiv:2001.08361","title":"Scaling Laws for Neural Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/82302458a6d03bc051b2e4fdb8277893a81d4bd6a029969045df6a8380904eac/"},{"source":"arxiv:2406.11813","target":"arxiv:2304.15004","evidence":"title","id":"arxiv:2304.15004","title":"Are Emergent Abilities of Large Language Models a Mirage?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d7cd8c081ad54f2fdaa2211a4b49dc21f6ff2db7282e287546d210ef90e84323/"},{"source":"arxiv:2406.11813","target":"arxiv:2309.14316","evidence":"title","id":"arxiv:2309.14316","title":"Physics of Language Models: Part 3.1, Knowledge Storage and Extraction","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b87376d90e99d2bb590e1a02ed50a55c90881591ed7d3e00c3d5be2aaacbb2af/"},{"source":"arxiv:2406.11813","target":"arxiv:2309.14402","evidence":"title","id":"arxiv:2309.14402","title":"Physics of Language Models: Part 3.2, Knowledge Manipulation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/79b4d0b92a11dcf3b111e9d2d1b6bb3f516bebce7333326f65e6e583d6ff3cf2/"}],"incoming":[{"source":"arxiv:2412.14135","target":"arxiv:2406.11813","evidence":"title","id":"arxiv:2412.14135","title":"Scaling of Search and Learning: A Roadmap to Reproduce o1 from Reinforcement Learning Perspective","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5a0b4b0b524698ee5f30eab659b7cd5104f5fe6497c8dec2d0c4b9d332350047/"},{"source":"arxiv:2501.00656","target":"arxiv:2406.11813","evidence":"id","id":"arxiv:2501.00656","title":"2 OLMo 2 Furious","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1ca88dba5d5c59e8cb23b2a304e74370b8ec69077b714a30d426300f7ef2a0b9/"},{"source":"arxiv:2503.21676","target":"arxiv:2406.11813","evidence":"title","id":"arxiv:2503.21676","title":"How do language models learn facts? Dynamics, curricula and hallucinations","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2ef59e7bcfcfc0e4bab638835504ad8a0d2f50665a9797edc09e8628e9b3c0b6/"}],"text_parts":[{"start":0,"end":15107,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html"},{"start":15107,"end":30506,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html"},{"start":30506,"end":46158,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html"},{"start":46158,"end":55187,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html"}],"contents":[{"title":"How Do Large Language Models Acquire Factual Knowledge During Pretraining?","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"Abstract","offset":78,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"1 Introduction","offset":1611,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"2 Related Work","offset":4894,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"3 Experimental Setup","offset":7198,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"Fictional Knowledge dataset","offset":7223,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"Evaluation metrics","offset":9122,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"Definition 1","offset":9979,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"Definition 2","offset":12149,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"Definition 3","offset":13048,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"Knowledge injection during pretraining","offset":15062,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-1.html#text"},{"title":"4 Results","offset":17008,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"4.1 Factual knowledge acquisition occurs by accumulating the observations of the fact","offset":17022,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"4.2 Effects of model scale and pretraining stage on knowledge acquisition dynamics","offset":19316,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"4.3 Forgetting in factual knowledge acquisition","offset":22676,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"Training steps and the forgetting of acquired factual knowledge have a power-law relationship","offset":22729,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"How quickly is the acquired factual knowledge lost?","offset":23764,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"Pretraining with a larger batch size helps LLMs acquire more knowledge","offset":25569,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"4.4 Implications for LLM pretraining","offset":27645,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"Why is popularity important for factual knowledge acquisition?","offset":27687,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"Why does deduplication enhance model performance?","offset":30450,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-2.html#text"},{"title":"5 Discussion and Conclusions","offset":31642,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"Acknowledgments and Disclosure of Funding","offset":33942,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"Appendix","offset":34375,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"Appendix A Limitations","offset":34388,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"Appendix B Dataset Construction and Examples","offset":35124,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"Appendix C Prompts Used for Dataset Generation","offset":37781,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"C.1 Prompts for the generation of injected knowledge","offset":37832,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"C.2 Prompts for the generation of paraphrased injected knowledge","offset":38929,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"C.3 Prompts for the generation of semantic generalization probes","offset":39186,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"C.4 Prompts for the generation of compositional generalization probes","offset":40557,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"C.5 Prompts for the validation of generated compositional generalization probes","offset":42543,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"Appendix D Detailed Training Setup","offset":46119,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-3.html#text"},{"title":"Appendix E Additional Figures for the Pretraining Experiments","offset":49928,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"E.1 Training dynamics of other OLMo-7B checkpoints","offset":49994,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"E.2 Training dynamics of other OLMo-1B checkpoints","offset":50050,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"E.3 Effectivity measurement data for OLMo-1B","offset":50106,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"E.4 Forgetting dynamics of OLMo-7B checkpoints","offset":50156,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"E.5 Forgetting dynamics of OLMo-1B checkpoints","offset":50208,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"Appendix F Experiments for Training Olmo-7B Checkpoints With a Constant Learning Rate","offset":50260,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"F.1 Training dynamics for constant learning rate experiments","offset":51586,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"Appendix G Forgetting Dynamics of Olmo-7B Trained With a Reduced Batch Size","offset":51652,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"G.1 Training dynamics for experiments on the forgetting dynamics with a reduced batch size","offset":53226,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"Appendix H Effect of the Number of Previous Encounters on Effectivity and Retainability of Factual Knowledge","offset":53322,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"},{"title":"Instructions for reporting errors","offset":54223,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/part-4.html#text"}],"text_code_points":55187,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a7b36ce24eb2182e7f0bba96c0cddcfe6efe610b830d5e8ad8aac2b7d14696c0/source.md"}