BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:A Data-Centric Approach to AI Adaptation and Alignment - Prof. Ste
 phen Bach (Brown University)
DTSTART:20260430T140000Z
DTEND:20260430T150000Z
UID:TALK246136@talks.cam.ac.uk
CONTACT:Lucas Resck
DESCRIPTION:Training generative AI is not a one-step process. In the case 
 of large language models (LLMs)\, self-supervision is often followed by su
 pervised and reinforcement learning stages to improve instruction followin
 g\, safety\, and other desirable qualities. This multi-stage process that 
 has emerged in the last 3 years has led to huge leaps in model capabilitie
 s. It has also led to new challenges and risks. In this talk\, I will over
 view some of our group's work to identify and address such challenges by f
 ocusing on the training data used at different stages. First\, I will disc
 uss the problem of adapting LLMs to new\, specialized domains and the role
  that synthetic\, i.e. LLM-generated\, training data can play. Then\, I wi
 ll share some of our work showing how mismatches in training data at diffe
 rent stages can lead to safety alignment risks. In one case\, LLMs with in
 adequate safety training can be more likely to respond to harmful queries 
 when presented in languages with less abundant data like Swahili or Scots 
 Gaelic. In another case\, LLMs trained to reason about solving math proble
 ms can then deploy those same reasoning skills to reason out of their own 
 safety guardrails. Together\, these findings highlight the importance of c
 areful training data management at all stages of AI development.\n\nBio: S
 tephen Bach is the Eliot Horowitz Assistant Professor of Computer Science 
 at Brown University. His latest research is on improving the processes by 
 which humans teach and instruct computers. That includes learning to gener
 alize from fewer examples\, with methods like zero-shot and few-shot learn
 ing\, as well as engineering training data\, with methods like synthetic d
 ata generation and programmatic weak supervision. He was a core contributo
 r to the Snorkel framework\, which was recognized with a Best of VLDB 2018
  award. Snorkel has been used in production at numerous Fortune 500 compan
 ies for programmatic training data curation. He also co-led the team that 
 developed the T0 family of large language models. The team was also one of
  the proposers of instruction tuning\, which is the process of fine-tuning
  language models with supervised training to follow instructions. Instruct
 ion tuning is now a standard part of training large language models. Steph
 en is also an advisor to Snorkel AI\, a company that provides software and
  services for data-centric AI.
LOCATION:https://cam-ac-uk.zoom.us/j/86890624365?pwd=oYGWpY7d5r3JOaUCaJXTD
 0sRECFxab.1
END:VEVENT
END:VCALENDAR