{"id":40225,"date":"2026-04-07T05:16:55","date_gmt":"2026-04-07T03:16:55","guid":{"rendered":"https:\/\/www.graviton.at\/letterswaplibrary\/fused-patent-arxiv-clustering-dataset-9m-raw-%e2%86%92-3-88m-release-bge-large-deterministic-quality-gating\/"},"modified":"2026-04-07T05:16:55","modified_gmt":"2026-04-07T03:16:55","slug":"fused-patent-arxiv-clustering-dataset-9m-raw-%e2%86%92-3-88m-release-bge-large-deterministic-quality-gating","status":"publish","type":"post","link":"https:\/\/www.graviton.at\/letterswaplibrary\/fused-patent-arxiv-clustering-dataset-9m-raw-%e2%86%92-3-88m-release-bge-large-deterministic-quality-gating\/","title":{"rendered":"Fused Patent + ArXiv Clustering Dataset (9M Raw \u2192 3.88M Release, BGE-large, Deterministic Quality Gating)"},"content":{"rendered":"<p><!-- SC_OFF --><\/p>\n<div class=\"md\">\n<p>Dataset link: <a href=\"https:\/\/huggingface.co\/datasets\/cjc0013\/ArvixFusedWithPatents\">https:\/\/huggingface.co\/datasets\/cjc0013\/ArvixFusedWithPatents<\/a> &lt;&#8212;&#8212;&#8212;&#8212; still uploading keep that in mind please<\/p>\n<p>9,063,272 raw rows \u2192 3,881,329 release rows (~20+ GB zipped)<\/p>\n<p>I built a zero-touch technical clustering pipeline over a fused patent + arXiv corpus. The full run was deterministic end-to-end, with Postgres used as the control plane rather than notebook state.<\/p>\n<p>This was not just \u201cembed some text and cluster it.\u201d<\/p>\n<p>The pipeline handled shard-level ingest\/normalization, chunk embeddings with <strong>BAAI\/bge-large-en-v1.5 (1024-dim)<\/strong>, clustering, reducer-tree merge, global assignment, BM25 artifact generation, and then a deterministic inspection\/gating pass to decide what was actually release-worthy.<\/p>\n<p>Full raw run output:<\/p>\n<ul>\n<li>91 label shards<\/li>\n<li>91 embedding shards<\/li>\n<li>91 chunk shards<\/li>\n<li>422 final clusters<\/li>\n<li>9,063,272 labeled rows<\/li>\n<\/ul>\n<p>I did <strong>not<\/strong> treat the raw output as valid by default.<\/p>\n<p>I ran deterministic inspection across all 422 clusters and split them into:<\/p>\n<ul>\n<li>147 coherent<\/li>\n<li>107 mixed<\/li>\n<li>168 metadata-heavy<\/li>\n<\/ul>\n<p>For the release dataset, I kept only the coherent clusters and dropped the mixed + metadata-heavy ones entirely.<\/p>\n<p>Final release subset:<\/p>\n<ul>\n<li>147 clusters<\/li>\n<li>3,881,329 rows<\/li>\n<li>42.82% retention from the raw run<\/li>\n<li>~20+ GB zipped<\/li>\n<\/ul>\n<p>I also generated deterministic cluster names from top terms as a lightweight inspection layer. Example release clusters looked like:<\/p>\n<ul>\n<li>wireless communications \/ device<\/li>\n<li>substrate \/ semiconductor \/ layer<\/li>\n<li>chemistry \/ formula \/ alkyl<\/li>\n<li>neural \/ data \/ network<\/li>\n<li>vehicle \/ system \/ control<\/li>\n<li>signal \/ data \/ circuit<\/li>\n<\/ul>\n<p>A big reason for the drop was metadata leakage. Some clusters were being driven by ingestion\/wrapper fields rather than actual technical content, so keeping everything would have made the dataset look cleaner than it really was.<\/p>\n<p>The system was also built to survive long, failure-prone runs instead of assuming ideal conditions. It uses Postgres-backed task leasing, heartbeats, and stage state; resumable progress; reducer-tree staged unblocking; explicit timeout handling; and a descending batch ladder so memory failures downshift deterministically instead of killing the run outright.<\/p>\n<p>I did not re-embed the corpus, hand-label clusters, manually patch results, or overwrite the original run. The release set is derived strictly from deterministic keep\/drop logic after full pipeline completion.<\/p>\n<p>The 147-cluster subset is the release-grade version.<\/p>\n<\/div>\n<p><!-- SC_ON -->   submitted by   <a href=\"https:\/\/www.reddit.com\/user\/Either_Pound1986\"> \/u\/Either_Pound1986 <\/a> <br \/> <span><a href=\"https:\/\/www.reddit.com\/r\/datasets\/comments\/1sej8ro\/fused_patent_arxiv_clustering_dataset_9m_raw_388m\/\">[link]<\/a><\/span>   <span><a href=\"https:\/\/www.reddit.com\/r\/datasets\/comments\/1sej8ro\/fused_patent_arxiv_clustering_dataset_9m_raw_388m\/\">[comments]<\/a><\/span><\/p><div class='watch-action'><div class='watch-position align-right'><div class='action-like'><a class='lbg-style1 like-40225 jlk' href='javascript:void(0)' data-task='like' data-post_id='40225' data-nonce='65e0e39b87' rel='nofollow'><img class='wti-pixel' src='https:\/\/www.graviton.at\/letterswaplibrary\/wp-content\/plugins\/wti-like-post\/images\/pixel.gif' title='Like' \/><span class='lc-40225 lc'>0<\/span><\/a><\/div><\/div> <div class='status-40225 status align-right'><\/div><\/div><div class='wti-clear'><\/div>","protected":false},"excerpt":{"rendered":"<p>Dataset link: https:\/\/huggingface.co\/datasets\/cjc0013\/ArvixFusedWithPatents &lt;&#8212;&#8212;&#8212;&#8212; still uploading keep that in mind please 9,063,272 raw rows \u2192 3,881,329 release&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[85],"tags":[],"class_list":["post-40225","post","type-post","status-publish","format-standard","hentry","category-datatards","wpcat-85-id"],"_links":{"self":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/posts\/40225","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/comments?post=40225"}],"version-history":[{"count":0,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/posts\/40225\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/media?parent=40225"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/categories?post=40225"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/tags?post=40225"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}