Author profile
Developer articles and tutorials by raihan-js.
Browse the latest writing surfaced through DevArt.
Frontier LLMs hallucinate FAR clause numbers somewhere between 0% and 32% of the time. A specialized 150M-parameter model trained in 4 minutes matches Claude Haiku on F1 with less than half the hallucination rate. Open dataset, open model, reproducible.
Raihan6 min read
Two months ago I started a portfolio project: build three small specialized language models for...
Raihan6 min read
Last month I published a 184M-parameter intent classifier that matches frontier LLMs at 22× lower...
Raihan10 min read
How a 184M-parameter DeBERTa fine-tune matches Claude Haiku 4.5 and GPT-4o within 4 points of accuracy on healthcare intent classification at 22× lower latency and ~$0 per inference.
Raihan9 min read