Workbench 04 · Four models, one headline
Give one financial headline to four real language models running on your device: FinBERT, trained on financial news; plain BERT, which has never seen a finance label; and two small chatbots, Qwen and Llama. Nothing is sent to a server.
Encoder · fine-tuned
BERT fine-tuned on financial news. Reads the whole headline at once and labels it positive, negative or neutral.
Loads on its own on computers. Pick a headline to see its call.
Encoder · general
The general-purpose original. It was only ever trained to fill in missing words, so we ask it to fill one in.
Loads on its own on computers. Pick a headline to see its guess.
Decoder · chat
A small chatbot. It writes its answer one token at a time, so you watch it think in real time.
Runs on your graphics chip. Downloads once, then your browser keeps it.
Decoder · chat
About twice Qwen3’s size. Compare how much a bigger model changes the answer.
Runs on your graphics chip. Downloads once, then your browser keeps it.
Encoders and decoders
BERT and FinBERT are encoders: they read the whole sentence at once and turn it into a judgement. Qwen and Llama are decoders: they predict the next word, over and over, which is what lets them write. Module 06 builds a tiny decoder from scratch.
Why fine-tuning matters
FinBERT is BERT with a few thousand labelled financial sentences added on top. That is enough to learn that “costs fell” is good news and “guidance cut” is bad, which general BERT often gets backwards. Try the restructuring headline.
Why the chatbots ramble
Qwen3 and Llama here are hundreds of times smaller than ChatGPT or Claude, squeezed to 4-bit numbers so they fit in your browser. They sound fluent but can be confidently wrong, which is exactly what module 14 is about.