Teaching a 9B Model to Investigate Production Alerts
datadog, Wednesday, September 23rd, 2026
Datadog fine-tuned a 9B-parameter Qwen3.5 model to investigate production alerts, matching most of GLM-5.3's accuracy far more cheaply.
Datadog describes fine-tuning Qwen3.5-9B into a specialized agent for change attribution, the task of identifying which recent change caused a production alert.
The resulting model achieved 87% of GLM-5.3's Recall@5 score while costing roughly 5% as much to run per investigation.
This suggests a small, purpose-tuned model can approach a much larger general-purpose model's accuracy on a narrow investigative task at a fraction of the inference cost, which Datadog frames as relevant for scaling automated alert investigation in Agent Observability.