​Hosted by Epidemic Sound - Search, Recs & Fine-Tuning: A Technical Evening

​Schema:
​17:30 — Doors open & Food
18:00 — Introduction
​18:10 — Gunay Abdullayeva @ Epidemic Sound
​18:40 — Break
18:55 — Aku Rouhe @ AMD Silo
19:25 — Break
19:30 — Anders Nyman @ Spotify
20:00 — Lightning Talk

​

Gunay Abdullayeva @ Epidemic Sound

​Generative Query Expansion: Fine-Tuning T5 for Real-Time Guided Search

​Summary: We leverage a fine-tuned T5 model to dynamically generate 'Guided Search' tags for user search queries. This talk covers the end-to-end engineering from curating high-quality training dataset based on historical user search experience to training and serving fine tuned T5 model.

​Aku Rouhe @ AMD Silo

​When Fine-Tuning Actually Matters

​Summary: Lessons from working with the SiloGen platform: where it worked well and where fine-tuning turned out to be relevant for only a small number of customers.

​We’ll focus on the practical question: when does it actually make sense to consider fine-tuning, and what should you try before going there?

​We’ll walk through common approaches such as supervised fine-tuning, DPO, RLHF, reinforcement learning from verifiable rewards, distillation, and continued pretraining. We’ll also touch on the Superficial Alignment Hypothesis and what happens when systems return to inference after tuning.

​Anders Nyman @ Spotify

​You Say Search, I Say Recs: A Scalable Agentic Approach to Query Understanding and Exploratory Search at Spotify

​Summary: Delivery a low latency solution for explicit intent understanding for Spotify Search bridging the gap between Recommendation and Search tasks. https://dl.acm.org/doi/10.1145/3705328.3748127

​#ai