News Gate

Open the demoCode
The News Gate results page: "Our 0.8B student beats its teacher." above bars for keyword rules, the untrained model, the GLM 5.3 teacher and the fine-tuned versions.

A fine-tuned 0.8B model that reads a stock headline, names the event type and decides whether it is a real shock or just noise.

Result
Beat its teacher, 90.5% to 79.3%
Event
GTC Berlin Builders Night with distil labs, Langfuse and Runpod
Role
Fine-tuned the model with LoRA
Built with
Qwen3.5-0.8B, LoRA, distil labs, Langfuse, Runpod
Team size
Missing, to be added

A market news feed produces tens of thousands of headlines, and almost all of them are noise. A trading system only needs to react to the few that change a company’s fundamentals: earnings, mergers, legal or regulatory action, management changes. Keyword rules miss most of them, and sending every headline to a frontier model is slow and expensive.

News Gate is Qwen3.5-0.8B fine-tuned with LoRA. It sorts each headline into one of ten event types in a single short call, then maps the type to shock or noise.

On the same 400 held-out headlines, 40 per label and never trained on, it gets the exact event type right 90.5% of the time. Its GLM 5.3 teacher gets 79.3%, the untuned base model 55.8% and keyword rules 30.5%.

distil labs ran the teacher, generated synthetic headlines and trained the student. Every call from the demo is traced in Langfuse, and a Langfuse experiment scores the deployed model, so misclassified headlines can become training rows for the next version.