BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York BEGIN:DAYLIGHT DTSTART:20260308T030000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:EDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20261101T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:EST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260427T233405Z UID:F15EF554-7274-4F4D-ACBF-5CCD3429545C DTSTART;TZID=America/New_York:20260625T190000 DTEND;TZID=America/New_York:20260625T200000 DESCRIPTION:This presentation outlines a practical pipeline for distilling large language models into compact\, in-house alternatives using mature te acher models. We focus on retrieval-augmented generation workflows\, where distilled students must learn context grounding\, citation discipline\, a nd latency-efficient behavior. The methodology covers data curation from t eacher-generated RAG trajectories\, constrained supervised fine-tuning wit h context masking\, and evaluation using faithfulness\, robustness\, and e fficiency metrics.\n\nWe demonstrate end-to-end integration with the Huggi ng Face platform\, leveraging Model Hub for teacher selection\, Datasets f or curated training data\, trl and peft for parameter-efficient training\, Spaces for interactive validation\, and Leaderboards for community benchm arking. Comparative analysis shows distillation achieves 85–95% of teach er performance at 20% inference cost\, enabling deployable\, privacy-compl iant RAG systems.\nThe talk concludes with mitigation strategies for commo n pitfalls and future directions in agentic and multi-teacher distillation .\n\nSpeaker(s): Zichao Li\, \n\nAgenda: \n7:00PM - Introduction of IEEE H amilton Section\n\n7:15PM - Presentation\n\n8:00PM - Q&A\n\n8:15PM - Refre shments\n\nRoom: Multipurpose Room 3\, Bldg: Trafalgar Park Community Cent re\, 133 Rebecca St\,\, Oakville\,\, Ontario\, Canada\, L6K 1J5 LOCATION:Room: Multipurpose Room 3\, Bldg: Trafalgar Park Community Centre\ , 133 Rebecca St\,\, Oakville\,\, Ontario\, Canada\, L6K 1J5 ORGANIZER:sneh@rchilli.com, eduardo.gomez.hennig@ieee.org SEQUENCE:23 SUMMARY:Distilling LLMs: Training an In-House Model with a Mature One for R etrieval-Augmented Generation URL;VALUE=URI:https://events.vtools.ieee.org/m/557483 X-ALT-DESC:Description: <br /><p class="MsoNormal">This presentation outlin es a practical pipeline for distilling large language models into compact\ , in-house alternatives using mature teacher models. We focus on retrieval -augmented generation workflows\, where distilled students must learn cont ext grounding\, citation discipline\, and latency-efficient behavior. The methodology covers data curation from teacher-generated RAG trajectories\, constrained supervised fine-tuning with context masking\, and evaluation using faithfulness\, robustness\, and efficiency metrics.<br><br>We demons trate end-to-end integration with the Hugging Face platform\, leveraging M odel Hub for teacher selection\, Datasets for curated training data\, trl and peft for parameter-efficient training\, Spaces for interactive validat ion\, and Leaderboards for community benchmarking. Comparative analysis sh ows distillation achieves 85&ndash\;95% of teacher performance at 20% infe rence cost\, enabling deployable\, privacy-compliant RAG systems. <br>The talk concludes with mitigation strategies for common pitfalls and future d irections in agentic and multi-teacher distillation.</p>\n<p><br><br></p>< br /><br />Agenda: <br /><p>7:00PM - Introduction of IEEE Hamilton Section </p>\n<p>7:15PM - Presentation</p>\n<p>8:00PM - Q&amp\;A</p>\n<p>8:15PM - Refreshments</p> END:VEVENT END:VCALENDAR