Spaces:
Running
Running
| title: Efficiency Lab | |
| emoji: ⚖️ | |
| colorFrom: blue | |
| colorTo: green | |
| sdk: static | |
| app_file: index.html | |
| pinned: false | |
| license: mit | |
| models: | |
| - ChadSaglam/swiss-bookkeeping-qwen-lora | |
| - ChadSaglam/swiss-bookkeeping-qwen-lora-mlx-8bit | |
| datasets: | |
| - ChadSaglam/swiss-sme-bookkeeping-dataset | |
| short_description: Accuracy vs memory vs speed for small bookkeeping models | |
| # Efficiency Lab | |
| Accuracy, latency and peak memory for five small bookkeeping models, measured on the same 74 held-out-vendor invoice lines with MLX on Apple Silicon. | |
| - [Fused model](https://huggingface.co/ChadSaglam/swiss-bookkeeping-qwen-lora) | |
| - [8-bit MLX model](https://huggingface.co/ChadSaglam/swiss-bookkeeping-qwen-lora-mlx-8bit) | |
| - [Dataset](https://huggingface.co/datasets/ChadSaglam/swiss-sme-bookkeeping-dataset) | |
| - [Benchmark code](https://github.com/ChadSaglam/efficiency-lab) | |
| The results come from synthetic data and one laptop. Small accuracy or latency differences should not be treated as a ranking. | |