{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2023:IQODKSZOHUWUXUMSB4TZ4JEQ56","short_pith_number":"pith:IQODKSZO","canonical_record":{"source":{"id":"2305.11206","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-05-18T17:45:22Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"14d423e617a7a71e863748e899f7d01d2437f1a915bbe2067b6f249d197c3682","abstract_canon_sha256":"036187de9440781b0da60b9626f55fce7779248ce12fd2c70b9cf39fce71bc51"},"schema_version":"1.0"},"canonical_sha256":"441c354b2e3d2d4bd1920f279e2490ef96f02f5009d25f2546417b376aa192e5","source":{"kind":"arxiv","id":"2305.11206","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2305.11206","created_at":"2026-05-17T23:38:14Z"},{"alias_kind":"arxiv_version","alias_value":"2305.11206v1","created_at":"2026-05-17T23:38:14Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2305.11206","created_at":"2026-05-17T23:38:14Z"},{"alias_kind":"pith_short_12","alias_value":"IQODKSZOHUWU","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"IQODKSZOHUWUXUMS","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"IQODKSZO","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2023:IQODKSZOHUWUXUMSB4TZ4JEQ56","target":"record","payload":{"canonical_record":{"source":{"id":"2305.11206","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-05-18T17:45:22Z","cross_cats_sorted":["cs.AI","cs.LG"],"title_canon_sha256":"14d423e617a7a71e863748e899f7d01d2437f1a915bbe2067b6f249d197c3682","abstract_canon_sha256":"036187de9440781b0da60b9626f55fce7779248ce12fd2c70b9cf39fce71bc51"},"schema_version":"1.0"},"canonical_sha256":"441c354b2e3d2d4bd1920f279e2490ef96f02f5009d25f2546417b376aa192e5","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:38:14.228937Z","signature_b64":"SdUP9o8+QqEqxFeppTCZquOBsCiyRDaPFZWahY/rENk+Epd71Eqn6YXxRN79/Px3YW7pxpzpOWc3S8osOK8mBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"441c354b2e3d2d4bd1920f279e2490ef96f02f5009d25f2546417b376aa192e5","last_reissued_at":"2026-05-17T23:38:14.228401Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:38:14.228401Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2305.11206","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:14Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"L9kpV1fczU36BEa7uN9Mv2DgB/Q7wZe0IePUGgTcamenOmFAVPO2fo2oko8FApltFVSArWkc1J7pNcFFmzehBg==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-30T19:23:49.735213Z"},"content_sha256":"a5823491113668f8c22a565c15081235ed5ffac7965040d146bdb7113d19f265","schema_version":"1.0","event_id":"sha256:a5823491113668f8c22a565c15081235ed5ffac7965040d146bdb7113d19f265"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2023:IQODKSZOHUWUXUMSB4TZ4JEQ56","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"LIMA: Less Is More for Alignment","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Large language models acquire nearly all knowledge during pretraining and need only limited curated data for alignment.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CL","authors_text":"Avia Efrat, Chunting Zhou, Gargi Ghosh, Jiao Sun, Lili Yu, Luke Zettlemoyer, Mike Lewis, Omer Levy, Pengfei Liu, Ping Yu, Puxin Xu, Srini Iyer, Susan Zhang, Xuezhe Ma, Yuning Mao","submitted_at":"2023-05-18T17:45:22Z","abstract_excerpt":"Large language models are trained in two stages: (1) unsupervised pretraining from raw text, to learn general-purpose representations, and (2) large scale instruction tuning and reinforcement learning, to better align to end tasks and user preferences. We measure the relative importance of these two stages by training LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow spec"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The 1,000 examples were selected and curated in a manner that captures the essential behaviors needed for broad generalization, and the human study protocol does not introduce systematic bias in preference judgments.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Fine-tuning a 65B model on 1,000 high-quality examples produces output that humans rate as good as or better than GPT-4 in 43% of cases, indicating most capabilities come from pretraining.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Large language models acquire nearly all knowledge during pretraining and need only limited curated data for alignment.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f8b7240721b0e8905a7b3a87964c19304f2a4738b09b2bd389aeb0fa03a02a08"},"source":{"id":"2305.11206","kind":"arxiv","version":1},"verdict":{"id":"d5b44e41-abaa-4924-8038-d1091cc18855","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T11:29:22.336718Z","strongest_claim":"Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.","one_line_summary":"Fine-tuning a 65B model on 1,000 high-quality examples produces output that humans rate as good as or better than GPT-4 in 43% of cases, indicating most capabilities come from pretraining.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The 1,000 examples were selected and curated in a manner that captures the essential behaviors needed for broad generalization, and the human study protocol does not introduce systematic bias in preference judgments.","pith_extraction_headline":"Large language models acquire nearly all knowledge during pretraining and need only limited curated data for alignment."},"references":{"count":43,"sample":[{"doi":"","year":null,"title":"Advances in Neural Information Processing Systems , volume=","work_id":"bd577a47-49ef-4f8a-81ca-86cd88b71479","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"The Tenth International Conference on Learning Representations , year=","work_id":"a74ea641-82bc-4498-abee-aca51256f9fe","ref_index":8,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.57967/hf/0513","year":null,"title":"doi:10.57967/hf/0513 , publisher =","work_id":"ad883f26-2d7a-434e-9ea7-1b9bbfbd31d6","ref_index":10,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"and Stoica, Ion and Xing, Eric P","work_id":"cb4b41f6-6d60-4db4-a4d1-6c5bb7899473","ref_index":11,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Hashimoto , title =","work_id":"fadc9cdf-612b-4081-85df-c1904ccef4c8","ref_index":12,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":43,"snapshot_sha256":"b81a57e929a79fdc64a19204e3767c3f11a5e40b0dd87280a7fcf66c79a75237","internal_anchors":7},"formal_canon":{"evidence_count":2,"snapshot_sha256":"b38186218baf3a2a486b307cd9a2d85b7fadba17c3d840bc3dbec5b1d605dc5d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"d5b44e41-abaa-4924-8038-d1091cc18855"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-17T23:38:14Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"ztChBbvc2orCg97lSPW+8YLpPoy9TviwsO89/Oeom5CmVps7sk8pWDPqGD+UiJKBxIci7cTO4s3blFn2EqZeDw==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-30T19:23:49.735761Z"},"content_sha256":"5690448459e958797e5725e371c63c2e2cc2e35198fb5e6493e1a3303427abe1","schema_version":"1.0","event_id":"sha256:5690448459e958797e5725e371c63c2e2cc2e35198fb5e6493e1a3303427abe1"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/IQODKSZOHUWUXUMSB4TZ4JEQ56/bundle.json","state_url":"https://pith.science/pith/IQODKSZOHUWUXUMSB4TZ4JEQ56/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/IQODKSZOHUWUXUMSB4TZ4JEQ56/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-30T19:23:49Z","links":{"resolver":"https://pith.science/pith/IQODKSZOHUWUXUMSB4TZ4JEQ56","bundle":"https://pith.science/pith/IQODKSZOHUWUXUMSB4TZ4JEQ56/bundle.json","state":"https://pith.science/pith/IQODKSZOHUWUXUMSB4TZ4JEQ56/state.json","well_known_bundle":"https://pith.science/.well-known/pith/IQODKSZOHUWUXUMSB4TZ4JEQ56/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2023:IQODKSZOHUWUXUMSB4TZ4JEQ56","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"036187de9440781b0da60b9626f55fce7779248ce12fd2c70b9cf39fce71bc51","cross_cats_sorted":["cs.AI","cs.LG"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-05-18T17:45:22Z","title_canon_sha256":"14d423e617a7a71e863748e899f7d01d2437f1a915bbe2067b6f249d197c3682"},"schema_version":"1.0","source":{"id":"2305.11206","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2305.11206","created_at":"2026-05-17T23:38:14Z"},{"alias_kind":"arxiv_version","alias_value":"2305.11206v1","created_at":"2026-05-17T23:38:14Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2305.11206","created_at":"2026-05-17T23:38:14Z"},{"alias_kind":"pith_short_12","alias_value":"IQODKSZOHUWU","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"IQODKSZOHUWUXUMS","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"IQODKSZO","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:5690448459e958797e5725e371c63c2e2cc2e35198fb5e6493e1a3303427abe1","target":"graph","created_at":"2026-05-17T23:38:14Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"The 1,000 examples were selected and curated in a manner that captures the essential behaviors needed for broad generalization, and the human study protocol does not introduce systematic bias in preference judgments."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Fine-tuning a 65B model on 1,000 high-quality examples produces output that humans rate as good as or better than GPT-4 in 43% of cases, indicating most capabilities come from pretraining."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Large language models acquire nearly all knowledge during pretraining and need only limited curated data for alignment."}],"snapshot_sha256":"f8b7240721b0e8905a7b3a87964c19304f2a4738b09b2bd389aeb0fa03a02a08"},"formal_canon":{"evidence_count":2,"snapshot_sha256":"b38186218baf3a2a486b307cd9a2d85b7fadba17c3d840bc3dbec5b1d605dc5d"},"paper":{"abstract_excerpt":"Large language models are trained in two stages: (1) unsupervised pretraining from raw text, to learn general-purpose representations, and (2) large scale instruction tuning and reinforcement learning, to better align to end tasks and user preferences. We measure the relative importance of these two stages by training LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow spec","authors_text":"Avia Efrat, Chunting Zhou, Gargi Ghosh, Jiao Sun, Lili Yu, Luke Zettlemoyer, Mike Lewis, Omer Levy, Pengfei Liu, Ping Yu, Puxin Xu, Srini Iyer, Susan Zhang, Xuezhe Ma, Yuning Mao","cross_cats":["cs.AI","cs.LG"],"headline":"Large language models acquire nearly all knowledge during pretraining and need only limited curated data for alignment.","license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-05-18T17:45:22Z","title":"LIMA: Less Is More for Alignment"},"references":{"count":43,"internal_anchors":7,"resolved_work":43,"sample":[{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":1,"title":"Advances in Neural Information Processing Systems , volume=","work_id":"bd577a47-49ef-4f8a-81ca-86cd88b71479","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":8,"title":"The Tenth International Conference on Learning Representations , year=","work_id":"a74ea641-82bc-4498-abee-aca51256f9fe","year":null},{"cited_arxiv_id":"","doi":"10.57967/hf/0513","is_internal_anchor":false,"ref_index":10,"title":"doi:10.57967/hf/0513 , publisher =","work_id":"ad883f26-2d7a-434e-9ea7-1b9bbfbd31d6","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":11,"title":"and Stoica, Ion and Xing, Eric P","work_id":"cb4b41f6-6d60-4db4-a4d1-6c5bb7899473","year":null},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":12,"title":"Hashimoto , title =","work_id":"fadc9cdf-612b-4081-85df-c1904ccef4c8","year":2023}],"snapshot_sha256":"b81a57e929a79fdc64a19204e3767c3f11a5e40b0dd87280a7fcf66c79a75237"},"source":{"id":"2305.11206","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-17T11:29:22.336718Z","id":"d5b44e41-abaa-4924-8038-d1091cc18855","model_set":{"reader":"grok-4.3"},"one_line_summary":"Fine-tuning a 65B model on 1,000 high-quality examples produces output that humans rate as good as or better than GPT-4 in 43% of cases, indicating most capabilities come from pretraining.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Large language models acquire nearly all knowledge during pretraining and need only limited curated data for alignment.","strongest_claim":"Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.","weakest_assumption":"The 1,000 examples were selected and curated in a manner that captures the essential behaviors needed for broad generalization, and the human study protocol does not introduce systematic bias in preference judgments."}},"verdict_id":"d5b44e41-abaa-4924-8038-d1091cc18855"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:a5823491113668f8c22a565c15081235ed5ffac7965040d146bdb7113d19f265","target":"record","created_at":"2026-05-17T23:38:14Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"036187de9440781b0da60b9626f55fce7779248ce12fd2c70b9cf39fce71bc51","cross_cats_sorted":["cs.AI","cs.LG"],"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2023-05-18T17:45:22Z","title_canon_sha256":"14d423e617a7a71e863748e899f7d01d2437f1a915bbe2067b6f249d197c3682"},"schema_version":"1.0","source":{"id":"2305.11206","kind":"arxiv","version":1}},"canonical_sha256":"441c354b2e3d2d4bd1920f279e2490ef96f02f5009d25f2546417b376aa192e5","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"441c354b2e3d2d4bd1920f279e2490ef96f02f5009d25f2546417b376aa192e5","first_computed_at":"2026-05-17T23:38:14.228401Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-17T23:38:14.228401Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"SdUP9o8+QqEqxFeppTCZquOBsCiyRDaPFZWahY/rENk+Epd71Eqn6YXxRN79/Px3YW7pxpzpOWc3S8osOK8mBg==","signature_status":"signed_v1","signed_at":"2026-05-17T23:38:14.228937Z","signed_message":"canonical_sha256_bytes"},"source_id":"2305.11206","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:a5823491113668f8c22a565c15081235ed5ffac7965040d146bdb7113d19f265","sha256:5690448459e958797e5725e371c63c2e2cc2e35198fb5e6493e1a3303427abe1"],"state_sha256":"3705f416f077f0c61ee3d12f96a4879b8aa9a56e2fdb61df15795dfbe5be9985"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"M3XgF9ijiTyXcAjDdiWxKaC1xMVZbDnKCbJSZ5s4SEA7shA2YeQIy9XoMgxX4w83NzHys80asI88CFAjIpiHBw==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-30T19:23:49.738236Z","bundle_sha256":"2adc4e7aea59957786c788abf887de12c3ec98752b7cfa79062f459710ac0dcc"}}