{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:AWBWG7T2LTPT5NTGNFJKGBBKTG","short_pith_number":"pith:AWBWG7T2","schema_version":"1.0","canonical_sha256":"0583637e7a5cdf3eb6666952a3042a99b5f508b05fe35d8ac191b6f1a665396a","source":{"kind":"arxiv","id":"2605.04948","version":2},"attestation_state":"computed","paper":{"title":"Adapting Large Language Models to a Low-Resource Agglutinative Language: A Comparative Study of LoRA and QLoRA for Bashkir","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"QLoRA on Mistral-7B and Phi-2 reaches near full fine-tuning perplexity for Bashkir with over 40 times fewer trainable parameters.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Mullosharaf K. Arabov, Svetlana S. Khaybullina","submitted_at":"2026-05-06T14:14:35Z","abstract_excerpt":"This paper presents a comparative study of parameter-efficient fine-tuning (PEFT) methods, including LoRA and QLoRA, applied to the task of adapting large language models to the Bashkir language, a low-resource agglutinative language of the Turkic family. Experimental evaluation is conducted on a Bashkir text corpus of 71k documents (46.9M tokens) using models of various architectures: DistilGPT2, GPT-2 (base, medium), Phi-2, Qwen2.5-7B, DeepSeek-7B, and Mistral-7B. To improve the reliability of results, each configuration was trained with three different random seeds.\n  The lowest perplexity "},"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":"2605.04948","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CL","submitted_at":"2026-05-06T14:14:35Z","cross_cats_sorted":[],"title_canon_sha256":"35df3c614deb2de9e8d5bcdbb2c94648e245f7d69825301ca81b82508341b446","abstract_canon_sha256":"42bfc4c333f2ae086c17dbfae01adefb3b03da57166a538f8b573bd63b08a958"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-02T03:05:05.623888Z","signature_b64":"JdBtIrlM9QOzKuwHkBGTn6bso8vFOIQVHkJL03XEt4d3SlkfLV58+ZSW24ECSLlf2zeAVtUrTFaF/C1mdrzMAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0583637e7a5cdf3eb6666952a3042a99b5f508b05fe35d8ac191b6f1a665396a","last_reissued_at":"2026-06-02T03:05:05.623385Z","signature_status":"signed_v1","first_computed_at":"2026-06-02T03:05:05.623385Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Adapting Large Language Models to a Low-Resource Agglutinative Language: A Comparative Study of LoRA and QLoRA for Bashkir","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"QLoRA on Mistral-7B and Phi-2 reaches near full fine-tuning perplexity for Bashkir with over 40 times fewer trainable parameters.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Mullosharaf K. Arabov, Svetlana S. Khaybullina","submitted_at":"2026-05-06T14:14:35Z","abstract_excerpt":"This paper presents a comparative study of parameter-efficient fine-tuning (PEFT) methods, including LoRA and QLoRA, applied to the task of adapting large language models to the Bashkir language, a low-resource agglutinative language of the Turkic family. Experimental evaluation is conducted on a Bashkir text corpus of 71k documents (46.9M tokens) using models of various architectures: DistilGPT2, GPT-2 (base, medium), Phi-2, Qwen2.5-7B, DeepSeek-7B, and Mistral-7B. To improve the reliability of results, each configuration was trained with three different random seeds.\n  The lowest perplexity "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"QLoRA applied to Mistral-7B (3.79) and Phi-2 (3.81) achieved comparable quality with over 40 times fewer trainable parameters.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the 71k-document Bashkir corpus is sufficiently representative for both training and reliable test-set evaluation, and that tokenizer compatibility issues do not dominate the observed performance differences across architectures.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"QLoRA on 7B-scale models like Mistral achieves perplexity within 0.45 of full fine-tuning on GPT-2 medium for Bashkir while using over 40 times fewer trainable parameters, though best perplexity does not guarantee coherent monolingual generation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"QLoRA on Mistral-7B and Phi-2 reaches near full fine-tuning perplexity for Bashkir with over 40 times fewer trainable parameters.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"46943ad90172ea4a60818ff0e8f103566e727a335fa8e6c235f8b28f16873923"},"source":{"id":"2605.04948","kind":"arxiv","version":2},"verdict":{"id":"152a888a-7329-433d-ba19-df3d6e51b91f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T16:18:49.365289Z","strongest_claim":"QLoRA applied to Mistral-7B (3.79) and Phi-2 (3.81) achieved comparable quality with over 40 times fewer trainable parameters.","one_line_summary":"QLoRA on 7B-scale models like Mistral achieves perplexity within 0.45 of full fine-tuning on GPT-2 medium for Bashkir while using over 40 times fewer trainable parameters, though best perplexity does not guarantee coherent monolingual generation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the 71k-document Bashkir corpus is sufficiently representative for both training and reliable test-set evaluation, and that tokenizer compatibility issues do not dominate the observed performance differences across architectures.","pith_extraction_headline":"QLoRA on Mistral-7B and Phi-2 reaches near full fine-tuning perplexity for Bashkir with over 40 times fewer trainable parameters."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.04948/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-20T10:40:00.944081Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T21:31:19.942052Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T13:59:48.104064Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"377890cc0a23b33606347a331abcf43fb74b14ec8d7acbc819db724112913c01"},"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":"2605.04948","created_at":"2026-06-02T03:05:05.623441+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.04948v2","created_at":"2026-06-02T03:05:05.623441+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.04948","created_at":"2026-06-02T03:05:05.623441+00:00"},{"alias_kind":"pith_short_12","alias_value":"AWBWG7T2LTPT","created_at":"2026-06-02T03:05:05.623441+00:00"},{"alias_kind":"pith_short_16","alias_value":"AWBWG7T2LTPT5NTG","created_at":"2026-06-02T03:05:05.623441+00:00"},{"alias_kind":"pith_short_8","alias_value":"AWBWG7T2","created_at":"2026-06-02T03:05:05.623441+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG","json":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG.json","graph_json":"https://pith.science/api/pith-number/AWBWG7T2LTPT5NTGNFJKGBBKTG/graph.json","events_json":"https://pith.science/api/pith-number/AWBWG7T2LTPT5NTGNFJKGBBKTG/events.json","paper":"https://pith.science/paper/AWBWG7T2"},"agent_actions":{"view_html":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG","download_json":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG.json","view_paper":"https://pith.science/paper/AWBWG7T2","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.04948&json=true","fetch_graph":"https://pith.science/api/pith-number/AWBWG7T2LTPT5NTGNFJKGBBKTG/graph.json","fetch_events":"https://pith.science/api/pith-number/AWBWG7T2LTPT5NTGNFJKGBBKTG/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG/action/storage_attestation","attest_author":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG/action/author_attestation","sign_citation":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG/action/citation_signature","submit_replication":"https://pith.science/pith/AWBWG7T2LTPT5NTGNFJKGBBKTG/action/replication_record"}},"created_at":"2026-06-02T03:05:05.623441+00:00","updated_at":"2026-06-02T03:05:05.623441+00:00"}