Alpamys Makazhan builds Soup, an open-source command-line workflow for fine-tuning and post-training language models. His work connects training configuration, memory management and checks that help decide whether a model should be shipped.[1][2]
Start with Soup →
Education
Links
GitHub Project site 
X · JustAlpamys Background
Contributions
Keep the frozen model outside GPU memory
Soup’s optional layer-streaming path keeps the frozen base model in CPU memory or on disk and transfers one decoder layer at a time to the GPU. Low-rank adapters remain trainable. This trades transfer overhead for a smaller memory requirement. The project marks the feature as beta and documents the hardware, configuration and correctness repairs behind its measurements.[2]
Make configuration and evaluation part of the workflow
Soup uses a YAML configuration to connect data, model, training method and output. Its workflow includes data inspection, configuration checks, training and export. The project’s site also describes evaluating saved checkpoints against expected behavior and recording a shipping decision alongside the weights. A successful training run alone does not establish that a model preserved the capabilities its user needs.[3]
Ideas
The repository distinguishes the lightweight command-line install from optional training dependencies. It documents unsupported settings and circumstances in which a memory estimate is only a warning. Its streaming measurements specify model, quantization and hardware instead of promising that every model will train on the same small GPU.[2]
Tools, writing & talks
Tools
Start here
Soupgithub.comA configurable CLI for data preparation, language-model post-training, evaluation and export.
Begin with the lightweight installation and configuration example before adding training dependencies. Actual fine-tuning needs compatible hardware and a suitable dataset; read the beta layer-streaming limits before planning around a small GPU.
Creator and maintainer[1][2]Original source
Writing, research & talks
- Soup project guidetrysoup.dev
A walkthrough of training, evaluation gates, layer streaming and the project’s stated limitations.