The WebMiner technical report presents our approach to extracting structured business data from HTML and Markdown, covering the progression from direct SFT to stepwise supervision and five-task mixed training, along with model evaluation results.
The horizontal axis shows cost per 100 examples (USD, log scale); the vertical axis shows offline scores. API costs are recorded charges, and local model costs are estimates for the generation stage.
The report text and figures are licensed under CC BY 4.0. Website implementation code is excluded from this license. Model weights and training/inference code remain closed-source and are also excluded.