The third post in my series on describing a classifier with Jev versus training one. This time I fine-tuned BERT on the same Jigsaw toxicity comments. With enough labels it ranks toxic comments better than Jev and the cloud-embedding models, but it needs about 500 labels to pass Jev, a GPU to train, and it misses almost every comment on the two rarest labels.
In my last post, I skipped training a spam classifier and used Jev instead. Here, I compare Jev with a neural network and k-NN built from thousands of labeled comments in the Jigsaw toxicity dataset: how accurate each one is, how many labels they need, and what they cost to run.
Quickly count tokens in Large Language Models (LLMs) like Gemini and Gemma using Rust. This efficient method avoids slow network calls, leveraging the `tokenizers` crate for local processing. The code example demonstrates token counting with minimal dependencies, even handling `aarch64` architecture challenges. Get started with fast, local token counting now!