Model Fine-tuning For Search: From Algorithms to Infra
summary
Fine-tuning turns pretrained encoders into searchable embeddings by cutting classification heads, attaching pooling layers and training with triplet/contrastive/cross-entropy or MSE losses; a pipeline uses a Taylor component to adapt PyTorch models, a trainer with hard-negative mining and cross-encoder pseudo-labeling, and a cloud orchestration via Argo workflows and autoscaling to spin up to 16 GPUs. Experiments use ResNet-52, T5 for query generation, CLIP-style dual encoders, and find 100 hard negatives (≈100k pairs) and a 20-minute training target give practical gains on English XMarket but need real user queries for non‑English datasets.
AI-written summary · about
chapters (6)
- 0:00Introduction to fine-tuning
- 2:28Motivation for model fine-tuning
- 7:48Embedding models and loss functions
- 19:35Scaling fine-tuning with cloud infra
- 23:33Data preparation and synthesis
- 30:36Conclusion and Q&A
related talks
41mFine-tuning Neural Sparse Model for Domain Specific Data From an Existing OpenSearch Index
33mNeural Search - Let's talk about quality
33mAI-Powered Search in Modern E-Commerce
28m🧼 From GPU-poor to data-rich: data quality practices for LLM fine-tuning
25mSemantic Search
10mFrom Pretrained To Personal: Privacy-First...
33mBuild Better AI with First-Party Data and Fine-Tuning
39mAI-Powered Search in Modern E-Commerce