Fine-tuning Neural Sparse Model for Domain Specific Data from...
summary
Fine-tunes an off‑the‑shelf neural sparse encoder on data pulled from an OpenSearch index, generating 238 training pairs from 49 documents by asking qwen 2.57b instruct to create five positive queries per doc and mining one hard plus one soft negative; trains doc‑only sparse head (in‑batch negatives=15, epochs=3) on an M4 MacBook without GPU, exports as a torchscript package in a sparse‑aft Docker image, deploys to OpenSearch and measures a clear NDCG uplift using Search Relevancy Workbench.
AI-written summary · about
chapters (16)
- 0:00Fine-tuning neural sparse models
- 4:41Generic sparse model challenges
- 7:50Solution and benefits
- 9:20Sparse encoder components
- 11:06Fine-tuning pipeline
- 13:14Demo setup and data prep
- 15:23Training data with LLM queries
- 19:16Training process and hyperparams
- 22:30Fine-tuning on MacBook
- 24:14Probing and validating model
- 26:01Exporting for OpenSearch
- 27:03Deploying in OpenSearch
- 27:44Comparing model outputs
- 31:03Search relevance improvements
- 34:14Scaling training data
- 38:10Key takeaways
related talks
41mFine-tuning Neural Sparse Model for Domain Specific Data From an Existing OpenSearch Index
44mRediscover your keyword search: Expand, Enrich and Rewrite
45mRevolutionizing Search with Power of Neural Sparse and Dense Vectors
09mFine-tuning embeddings 4 vector search
43mSupercharging your transformers with synthetic query generation and lexical search
47mBoosting AI With OpenSearch: Vector Stores, Similarity Search, and Retrieval-Augme...
40mSparse models are fast models: Improving DNN inference performance by over 10X
28mUse OpenSearch as Your Vector Search Database To Optimize GenAI Accuracy!