{"paper":{"title":"LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"LLaMA-Adapter adapts frozen LLaMA to follow instructions using only 1.2 million added parameters.","cross_cats":["cs.AI","cs.CL","cs.LG","cs.MM"],"primary_cat":"cs.CV","authors_text":"Aojun Zhou, Chris Liu, Hongsheng Li, Jiaming Han, Pan Lu, Peng Gao, Renrui Zhang, Shilin Yan, Xiangfei Hu, Yu Qiao","submitted_at":"2023-03-28T17:59:12Z","abstract_excerpt":"We present LLaMA-Adapter, a lightweight adaption method to efficiently fine-tune LLaMA into an instruction-following model. Using 52K self-instruct demonstrations, LLaMA-Adapter only introduces 1.2M learnable parameters upon the frozen LLaMA 7B model, and costs less than one hour for fine-tuning on 8 A100 GPUs. Specifically, we adopt a set of learnable adaption prompts, and prepend them to the word tokens at higher transformer layers. Then, a zero-initialized attention mechanism with zero gating is proposed, which adaptively injects the new instructional cues into LLaMA, while effectively pres"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"With our efficient training, LLaMA-Adapter can generate high-quality responses, comparable to Alpaca with fully fine-tuned 7B parameters.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The zero-initialized attention mechanism with zero gating adaptively injects the new instructional cues into LLaMA while effectively preserving its pre-trained knowledge.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLaMA-Adapter turns frozen LLaMA 7B into a capable instruction follower using only 1.2M new parameters and zero-init attention, matching Alpaca while extending to image-conditioned reasoning on ScienceQA and COCO.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LLaMA-Adapter adapts frozen LLaMA to follow instructions using only 1.2 million added parameters.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"50a8db6213e450be59e73c21dc566ff19f998250c31b89e53264e3b522e0b4ee"},"source":{"id":"2303.16199","kind":"arxiv","version":3},"verdict":{"id":"033679df-4a70-4f12-9af8-1266f199bfda","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T23:01:25.060113Z","strongest_claim":"With our efficient training, LLaMA-Adapter can generate high-quality responses, comparable to Alpaca with fully fine-tuned 7B parameters.","one_line_summary":"LLaMA-Adapter turns frozen LLaMA 7B into a capable instruction follower using only 1.2M new parameters and zero-init attention, matching Alpaca while extending to image-conditioned reasoning on ScienceQA and COCO.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The zero-initialized attention mechanism with zero gating adaptively injects the new instructional cues into LLaMA while effectively preserving its pre-trained knowledge.","pith_extraction_headline":"LLaMA-Adapter adapts frozen LLaMA to follow instructions using only 1.2 million added parameters."},"references":{"count":278,"sample":[{"doi":"","year":2023,"title":"Alpaca-lora. https://github.com/tloen/alpaca-lora, 2023","work_id":"9a06075c-ae6a-4dd1-a289-ebde72983a0f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Flamingo: a visual language model for few-shot learning","work_id":"01d90a48-13d8-4bc8-a06a-97adb5201146","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Open llm leaderboard","work_id":"5d0d6bb6-6bf7-45f4-a39e-40222560248f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1901,"title":"Language models are few-shot learners","work_id":"b5af3a68-2622-4421-b39b-b1d2fbde2d8d","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2004,"title":"Introduction to the conll-2004 shared task: Semantic role labeling","work_id":"c0bd2045-d243-4c6d-a996-2634541c2b6b","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":278,"snapshot_sha256":"14c936933f9c610e90f0dbbeae5ab392bfc5944dc15cd86e833a0b401fe6c09b","internal_anchors":45},"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"}