{"id":42916,"date":"2026-10-09T15:27:10","date_gmt":"2026-10-09T13:27:10","guid":{"rendered":"https:\/\/www.graviton.at\/letterswaplibrary\/self-promotion-synthetic-support-ticket-routing-125-labelled-messages-plus-five-models-recorded-answers-cc-by-4-0\/"},"modified":"2026-10-09T15:27:10","modified_gmt":"2026-10-09T13:27:10","slug":"self-promotion-synthetic-support-ticket-routing-125-labelled-messages-plus-five-models-recorded-answers-cc-by-4-0","status":"publish","type":"post","link":"https:\/\/www.graviton.at\/letterswaplibrary\/self-promotion-synthetic-support-ticket-routing-125-labelled-messages-plus-five-models-recorded-answers-cc-by-4-0\/","title":{"rendered":"[self-promotion] [synthetic] Support-ticket Routing: 125 Labelled Messages Plus Five Models&#8217; Recorded Answers (CC BY 4.0)"},"content":{"rendered":"<p><!-- SC_OFF --><\/p>\n<div class=\"md\">\n<p>I&#8217;ve published a small text-classification dataset for support-ticket routing, together with the per-case answers of five models. It&#8217;s my own project.<\/p>\n<p><strong>What&#8217;s in it<\/strong><\/p>\n<ul>\n<li><strong>125 fictional English support messages<\/strong>, each labelled with one of five teams: <code>billing<\/code>, <code>technical<\/code>, <code>account<\/code>, <code>sales<\/code>, or <code>needs_clarification<\/code> (the message doesn&#8217;t say enough to route).<\/li>\n<li><strong>Splits:<\/strong> 25 development cases and 100 held-out test cases, balanced at 25 per label overall.<\/li>\n<li><strong>Per case:<\/strong> the expected label, a written rationale, tags, a category (clear, boundary, ambiguous, adversarial) and a difficulty.<\/li>\n<li><strong>The label policy<\/strong> as a separate file, including precedence rules such as &#8220;route the immediate blocker, not the eventual goal&#8221;.<\/li>\n<li><strong>500 result rows:<\/strong> Jev, Clef, Clef Flash, OpenAI&#8217;s Decisions API and GPT-6 Luna with structured output on the 100 test cases, with each answer, status, response time and token counts, and per-label probabilities where the provider returns them.<\/li>\n<\/ul>\n<p><strong>How it was made<\/strong><\/p>\n<p>The messages are fictional, drafted with AI assistance under the written policy, then reviewed by two people. No real customer conversations were used.<\/p>\n<p><strong>Limits<\/strong><\/p>\n<ul>\n<li>It is small and English only.<\/li>\n<li>It is balanced by design, so it is not representative of a real queue.<\/li>\n<li>The test cases are now public, so they are no longer a clean held-out set for future tuning.<\/li>\n<\/ul>\n<p><strong>Links<\/strong><\/p>\n<ul>\n<li>Dataset (CC BY 4.0): <a href=\"https:\/\/huggingface.co\/datasets\/decisionmodelhub\/support-routing\">huggingface.co\/datasets\/decisionmodelhub\/support-routing<\/a><\/li>\n<li>Results page with a case-by-case explorer: <a href=\"https:\/\/decisionmodelhub.com\/benchmarks\/support-routing\">decisionmodelhub.com\/benchmarks\/support-routing<\/a><\/li>\n<\/ul>\n<p><strong>Load it<\/strong><\/p>\n<pre><code>from datasets import load_dataset cases = load_dataset(\"decisionmodelhub\/support-routing\", \"cases\") results = load_dataset(\"decisionmodelhub\/support-routing\", \"results\", split=\"test\") <\/code><\/pre>\n<p>If you work with support data: what would make a larger version more useful to you? Longer messages, realistic label imbalance, more languages?<\/p>\n<\/div>\n<p><!-- SC_ON -->   submitted by   <a href=\"https:\/\/www.reddit.com\/user\/AncientComment2352\"> \/u\/AncientComment2352 <\/a> <br \/> <span><a href=\"https:\/\/www.reddit.com\/r\/datasets\/comments\/1x1ke7w\/selfpromotion_synthetic_supportticket_routing_125\/\">[link]<\/a><\/span>   <span><a href=\"https:\/\/www.reddit.com\/r\/datasets\/comments\/1x1ke7w\/selfpromotion_synthetic_supportticket_routing_125\/\">[comments]<\/a><\/span><\/p><div class='watch-action'><div class='watch-position align-right'><div class='action-like'><a class='lbg-style1 like-42916 jlk' href='javascript:void(0)' data-task='like' data-post_id='42916' data-nonce='b9a40cecc8' 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-42916 lc'>0<\/span><\/a><\/div><\/div> <div class='status-42916 status align-right'><\/div><\/div><div class='wti-clear'><\/div>","protected":false},"excerpt":{"rendered":"<p>I&#8217;ve published a small text-classification dataset for support-ticket routing, together with the per-case answers of five models&#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-42916","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\/42916","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=42916"}],"version-history":[{"count":0,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/posts\/42916\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/media?parent=42916"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/categories?post=42916"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.graviton.at\/letterswaplibrary\/wp-json\/wp\/v2\/tags?post=42916"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}