{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:PQUDFZBHSYQ6SKKVVYG5LXURLV","short_pith_number":"pith:PQUDFZBH","schema_version":"1.0","canonical_sha256":"7c2832e4279621e92955ae0dd5de915d488bd9926869152e31734a65d25e3f1c","source":{"kind":"arxiv","id":"2405.13155","version":1},"attestation_state":"computed","paper":{"title":"ReALLM: A general framework for LLM compression and fine-tuning","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Eric Moulines, Lisa Bedin, Louis Leconte, Van Minh Nguyen","submitted_at":"2024-05-21T18:50:51Z","abstract_excerpt":"We introduce ReALLM, a novel approach for compression and memory-efficient adaptation of pre-trained language models that encompasses most of the post-training quantization and fine-tuning methods for a budget of <4 bits. Pre-trained matrices are decomposed into a high-precision low-rank component and a vector-quantized latent representation (using an autoencoder). During the fine-tuning step, only the low-rank components are updated. Our results show that pre-trained matrices exhibit different patterns. ReALLM adapts the shape of the encoder (small/large embedding, high/low bit VQ, etc.) to e"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2405.13155","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","primary_cat":"cs.LG","submitted_at":"2024-05-21T18:50:51Z","cross_cats_sorted":[],"title_canon_sha256":"2446a8d112dbea4a46334361aa52fed25ca9acaa13797b69db39f187cc657bb2","abstract_canon_sha256":"f7e38fc06c5ba08d3028197252beaa598b3076b5846693705ac574b42a7bef53"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:21:58.715466Z","signature_b64":"cxN2naMM+zxvPRuMk++YCGBI5UfpV8O0gscLw9lBjMp7kLA3W91SvRA2LcmVsrlYTS3GY2MrCKCUtSED4p5NAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"7c2832e4279621e92955ae0dd5de915d488bd9926869152e31734a65d25e3f1c","last_reissued_at":"2026-07-05T08:21:58.715003Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:21:58.715003Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"ReALLM: A general framework for LLM compression and fine-tuning","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.LG","authors_text":"Eric Moulines, Lisa Bedin, Louis Leconte, Van Minh Nguyen","submitted_at":"2024-05-21T18:50:51Z","abstract_excerpt":"We introduce ReALLM, a novel approach for compression and memory-efficient adaptation of pre-trained language models that encompasses most of the post-training quantization and fine-tuning methods for a budget of <4 bits. Pre-trained matrices are decomposed into a high-precision low-rank component and a vector-quantized latent representation (using an autoencoder). During the fine-tuning step, only the low-rank components are updated. Our results show that pre-trained matrices exhibit different patterns. ReALLM adapts the shape of the encoder (small/large embedding, high/low bit VQ, etc.) to e"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2405.13155","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2405.13155/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2405.13155","created_at":"2026-07-05T08:21:58.715069+00:00"},{"alias_kind":"arxiv_version","alias_value":"2405.13155v1","created_at":"2026-07-05T08:21:58.715069+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2405.13155","created_at":"2026-07-05T08:21:58.715069+00:00"},{"alias_kind":"pith_short_12","alias_value":"PQUDFZBHSYQ6","created_at":"2026-07-05T08:21:58.715069+00:00"},{"alias_kind":"pith_short_16","alias_value":"PQUDFZBHSYQ6SKKV","created_at":"2026-07-05T08:21:58.715069+00:00"},{"alias_kind":"pith_short_8","alias_value":"PQUDFZBH","created_at":"2026-07-05T08:21:58.715069+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":1,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2412.01129","citing_title":"RILQ: Rank-Insensitive LoRA-based Quantization Error Compensation for Boosting 2-bit Large Language Model Accuracy","ref_index":33,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV","json":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV.json","graph_json":"https://pith.science/api/pith-number/PQUDFZBHSYQ6SKKVVYG5LXURLV/graph.json","events_json":"https://pith.science/api/pith-number/PQUDFZBHSYQ6SKKVVYG5LXURLV/events.json","paper":"https://pith.science/paper/PQUDFZBH"},"agent_actions":{"view_html":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV","download_json":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV.json","view_paper":"https://pith.science/paper/PQUDFZBH","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2405.13155&json=true","fetch_graph":"https://pith.science/api/pith-number/PQUDFZBHSYQ6SKKVVYG5LXURLV/graph.json","fetch_events":"https://pith.science/api/pith-number/PQUDFZBHSYQ6SKKVVYG5LXURLV/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV/action/timestamp_anchor","attest_storage":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV/action/storage_attestation","attest_author":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV/action/author_attestation","sign_citation":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV/action/citation_signature","submit_replication":"https://pith.science/pith/PQUDFZBHSYQ6SKKVVYG5LXURLV/action/replication_record"}},"created_at":"2026-07-05T08:21:58.715069+00:00","updated_at":"2026-07-05T08:21:58.715069+00:00"}