News
We'd rather show the work than describe it: reproducible evaluations on public data, harness open-sourced, every number checkable.
A reproducible four-way ablation of Apertus-8B on Swiss data protection Q&A over the Federal Act on Data Protection (revised FADP / nDSG) (SR 235.1). Fine-tuning plus retrieval takes the production bar (answer + exact citation) from 37.0% to 54.3% — deterministic scoring, public Fedlex data, harness open-sourced. Industry: Government & Public Sector.
Read the reportA reproducible four-way ablation of Apertus-8B on Swiss therapeutic-products regulation Q&A over the Therapeutic Products Act (Heilmittelgesetz) (SR 812.21). Fine-tuning plus retrieval takes the production bar (answer + exact citation) from 22.7% to 68.2% — deterministic scoring, public Fedlex data, harness open-sourced. Industry: Healthcare & Pharma.
Read the reportA second reproducible ablation, this time on the Swiss Code of Obligations. Fine-tuning + retrieval wins at 52.3% — but only after we caught a training mistake that made the naive fine-tune worse than retrieval alone. The failure, diagnosed and fixed.
Read the reportA reproducible four-way ablation of Apertus-8B on Swiss product-safety law Q&A over the Product Safety Act (Produktesicherheitsgesetz) (SR 930.11). Fine-tuning plus retrieval takes the production bar (answer + exact citation) from 45.0% to 60.0% — deterministic scoring, public Fedlex data, harness open-sourced. Industry: Manufacturing & Industrial.
Read the reportA reproducible four-way ablation of Apertus-8B on Swiss insurance-contract law Q&A over the Insurance Contract Act (Versicherungsvertragsgesetz) (SR 221.229.1). Fine-tuning plus retrieval takes the production bar (answer + exact citation) from 18.8% to 37.5% — deterministic scoring, public Fedlex data, harness open-sourced. Industry: Insurance.
Read the reportA reproducible four-way ablation of Apertus-8B on Swiss FINMA circular Q&A. Fine-tuning plus retrieval doubles the production bar (answer + exact citation) from 22.4% to 45.4% — scored deterministically, on public data, with the harness open-sourced.
Read the reportThe reports here run on public data. In a Pilot the same method runs on yours — on your documents, in your jurisdiction.