SDS Knowledge Base — RAG + LoRA Fine-Tuning Built a local question-answering system over 497,000 ...SDS Knowledge Base — RAG + LoRA Fine-Tuning Built a local question-answering system over 497,000 ...
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
Henry's avatar
Henry Du
• Aug 5
SDS Knowledge Base — RAG + LoRA Fine-Tuning
Built a local question-answering system over 497,000 chemical safety data sheets: two parsers unify raw corpora (Kaggle SDS files + PubChem XML) into one validated schema, and a hybrid retrieval pipeline SQLite FTS5 for exact identifiers, vector embeddings for semantics, fused with reciprocal rank fusion and re-ranked by a cross-encoder answers queries with cited sources and honest "no data" responses instead of guesses. Fine-tuned Qwen 2.5 7B with LoRA on 4,000+ Q&A pairs generated from the corpus, exported the adapter to GGUF.
Vibe CodingPyTorch
Post image
Jasper's avatar
Jasper Rexford
pro
• Aug 5
RAG plus LoRA over safety data sheets is a domain where a confident wrong answer actually hurts someone, so the retrieval bar is higher than a normal knowledge base. Curious what the LoRA earns you on top of retrieval here, domain formatting and vocabulary, or recall the...
Henry's avatar
Henry Du
• Aug 5
Thank you for your reply. LoRA doesn’t really change retrieval, the search is identical with or without it. It does help the model to reduce hallucinations from base model training data, which I found is often wrong for rare chemicals.
Before the model is allowed to answer,...
Back to feed
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started