Engineering case study

A 47B mixture-of-experts model, fine-tuned and served

Mixtral-8x7B-v0.1 adapted with QLoRA on Databricks Dolly 15K, trained on AWS SageMaker, and served through API Gateway and Lambda behind this console.

Base modelMixtral-8x7B-v0.1
AdaptationQLoRA · rank 64
Quantization4-bit NF4
Trainable params~3.96%

Evidence

Watch the run finish1528/1528 steps · SUCCESS
The job that produced the adapterCompleted · approximately 7 hours
SageMaker training job list showing a completed QLoRA fine-tuning job.
The endpoint that serves itInService
SageMaker endpoint detail showing status InService.
Run inference against the deployed endpointFull engineering case study — problem, approach, architecture, decisions