{"paper":{"title":"Textbooks Are All You Need II: phi-1.5 technical report","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A 1.3 billion parameter model trained on synthetic textbooks matches models five times larger on reasoning tasks.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Allie Del Giorno, Ronen Eldan, S\\'ebastien Bubeck, Suriya Gunasekar, Yin Tat Lee, Yuanzhi Li","submitted_at":"2023-09-11T14:01:45Z","abstract_excerpt":"We continue the investigation into the power of smaller Transformer-based language models as initiated by \\textbf{TinyStories} -- a 10 million parameter model that can produce coherent English -- and the follow-up work on \\textbf{phi-1}, a 1.3 billion parameter model with Python coding performance close to the state-of-the-art. The latter work proposed to use existing Large Language Models (LLMs) to generate ``textbook quality\" data as a way to enhance the learning process compared to traditional web data. We follow the ``Textbooks Are All You Need\" approach, focusing this time on common sense"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"phi-1.5 ... with performance on natural language tasks comparable to models 5x larger, and surpassing most non-frontier LLMs on more complex reasoning tasks such as grade-school mathematics and basic coding.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That standard benchmarks for grade-school math and basic coding sufficiently measure general common sense reasoning without the model overfitting to patterns in the synthetic textbook data.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"phi-1.5 is a 1.3B parameter model trained on synthetic textbook data that matches the reasoning performance of models five times larger on natural language, math, and basic coding tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A 1.3 billion parameter model trained on synthetic textbooks matches models five times larger on reasoning tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"bd49c7239b4b89d801f193cb561cadc90abf68ae81b2ff13704a29d11ade9841"},"source":{"id":"2309.05463","kind":"arxiv","version":1},"verdict":{"id":"29c3349e-ead8-483e-b856-738cf676389a","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:13:13.944635Z","strongest_claim":"phi-1.5 ... with performance on natural language tasks comparable to models 5x larger, and surpassing most non-frontier LLMs on more complex reasoning tasks such as grade-school mathematics and basic coding.","one_line_summary":"phi-1.5 is a 1.3B parameter model trained on synthetic textbook data that matches the reasoning performance of models five times larger on natural language, math, and basic coding tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That standard benchmarks for grade-school math and basic coding sufficiently measure general common sense reasoning without the model overfitting to patterns in the synthetic textbook data.","pith_extraction_headline":"A 1.3 billion parameter model trained on synthetic textbooks matches models five times larger on reasoning tasks."},"references":{"count":24,"sample":[{"doi":"","year":null,"title":"Program Synthesis with Large Language Models","work_id":"fd241a05-03b9-4de2-9588-9d77ce176125","ref_index":1,"cited_arxiv_id":"2108.07732","is_internal_anchor":true},{"doi":"","year":null,"title":"Identify, align, and integrate: Matching knowledge graphs to commonsense reasoning tasks","work_id":"9f1ee709-7c10-4e2e-a368-5342f156dbfa","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Sparks of Artificial General Intelligence: Early experiments with GPT-4","work_id":"a23cfe92-7f7c-424b-98d4-b386a83002fb","ref_index":3,"cited_arxiv_id":"2303.12712","is_internal_anchor":true},{"doi":"","year":2021,"title":"On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency , pages 610–623","work_id":"c49bc9e9-444d-46e0-bdec-fc27a248705a","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1911,"title":"PIQA: Reasoning about Physical Commonsense in Natural Language","work_id":"0d865a62-6376-4606-8d3a-eeb3b6e9ba6d","ref_index":5,"cited_arxiv_id":"1911.11641","is_internal_anchor":true}],"resolved_work":24,"snapshot_sha256":"baa61333aed4ee2a2ffbf29b8d56e89cb69cf4cf24f0ae4bb428cc66c971777d","internal_anchors":14},"formal_canon":{"evidence_count":2,"snapshot_sha256":"5e6edb4a9c5139f8230b9008f46ba426ec5f32b7731d9a02315bf3ad3947f19c"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}