| title: README | |
| emoji: 🌍 | |
| colorFrom: yellow | |
| colorTo: gray | |
| sdk: static | |
| pinned: false | |
| license: apache-2.0 | |
| # unfork | |
| **unfork** is the home of the **AMOS decision-model family** — generative decision models that turn a single forward pass into a **normalized probability distribution**, served through the **systemone structured decision protocol**. | |
| ## What we build | |
| Instead of free-form chat answers, AMOS models emit fully structured decision primitives: | |
| | Primitive | Output | | |
| |---|---| | |
| | `choice` | A probability distribution over candidate options | | |
| | `score` | Calibrated numeric scores for each candidate | | |
| | `noul` | An explicit "no usable label" distribution for out-of-scope inputs | | |
| This design yields **100% parseable structured output**, calibrated probabilities, and low end-to-end latency — purpose-built for routing, classification, triage, and automated decision pipelines. | |
| ## Flagship release | |
| **[unfork/amos-2.8](https://huggingface.co/unfork/amos-2.8)** — a QLoRA adapter on **Qwen3-8B**, trained on 50,567 curriculum-based decision tasks. | |
| Highlights from independent randomized blind tests (rule-based gold scoring, zero text-level overlap with training data): | |
| - **500-question fresh blind board**: **96.6%** raw accuracy (amos-2.8) vs **96.4%** for jev-1.13 (commercial API) and **42.8%** for laya (zero-shot); **98.6%** on pure decision tasks; **96.05%** under strict out-of-distribution evaluation. | |
| - **6,000-question three-model same board**: **86.77%** on main tasks vs 72.28% (jev-1.13) and 45.07% (laya); **96.23%** on out-of-scope attributes vs 86.87% and 43.89%. | |
| - **0 API errors** across all boards; ≈0.9–1.2 s warm end-to-end latency through the public gateway. | |
| ## Links | |
| - Model: [unfork/amos-2.8](https://huggingface.co/unfork/amos-2.8) | |
| - License: Apache-2.0 | |