Air-Quality SLM v3
a compact language model for environmental monitoring
2 min readupdated
Air-Quality SLM v3 is a specialized small language model (SLM) featuring a Qwen3-style architecture with approximately 20.4 million parameters. The model is trained on urban air quality sensor data and comes in GGUF and PyTorch formats for local deployment. Version 3 introduces a streamlined Ollama template with system prompt support, building on the weights from the previous training stage. It is designed for developers and researchers exploring lightweight domain-specific models on modest hardware.
Key features
- Architecture: Custom Qwen3-style block with grouped-query attention and RMSNorm.
- Format: Available as GGUF (F16) and PyTorch state dict.
- Deployment: Full integration with Ollama via a ready-to-use Modelfile.
- Limitations: Context window of only 128 tokens and a narrow domain focus.
What you can do
- Run the model locally via Ollama for testing and integration.
- Test text generation outputs using the UCI Air Quality dataset.
- Load the model directly through a custom PyTorch class using the training script.
Example
# Запуск через Ollama из папки с GGUF и Modelfile
ollama create air-quality-slm-v3 -f Modelfile
ollama run air-quality-slm-v3Read more
DetailsBenchmarks (v2): Validation loss — 0.4877, test loss — 0.5555, test perplexity — 1.743.
Hardware requirements: Extremely lightweight model trained on a Ryzen 5 5600H without a GPU.
Limitations: High error rate in numerical calculations, sensor hallucinations, and low response diversity.
Training details: 614,400 tokens on the UCI Air Quality dataset (2004–2005).










Comments