| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3.8-27B |
| library_name: peft |
| language: |
| - en |
| tags: |
| - decision-making |
| - classification |
| - tool-selection |
| - lora |
| --- |
| # JADE |
|
|
| **Just Another Decision Engine.** |
|
|
| Give JADE some context, a question, and a set of options. It returns a decision |
| and a probability for each option. It can also answer yes/no questions with a |
| probability. There is no generated explanation to parse. |
|
|
| JADE is built on [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B), with |
| rank-8 LoRA adapters and a trained 255-way decision head. This release was |
| trained on 160,017 synthetic examples covering reasoning, language, |
| classification, retrieval, tools, and probabilistic decisions. |
|
|
| ## Using JADE |
|
|
| The included inference code uses Python 3.11+ and vLLM. You will need an NVIDIA GPU with enough |
| memory for the 27B base model and its inference cache. We tested this release |
| on a B200 with vLLM 0.30.0. |
|
|
| ```bash |
| pip install vllm==0.30.0 huggingface_hub |
| hf download theunnecessarythings/JADE --local-dir ./jade-model |
| export PYTHONPATH="$PWD/jade-model:$PYTHONPATH" |
| ``` |
|
|
| ```python |
| from jade import JadeEngine |
| |
| engine = JadeEngine(model="./jade-model") |
| |
| response, _ = engine( |
| state="Move my meeting with Sam to tomorrow at 3 pm.", |
| questions={ |
| "tool": { |
| "type": "choice", |
| "instructions": "Which tool should handle this request?", |
| "criteria": { |
| "calendar": "Find and update a calendar event", |
| "email": "Search the email inbox", |
| "notes": "Create a note", |
| }, |
| } |
| }, |
| ) |
| |
| print(response["answers"]["tool"]) |
| # Contains "choice" and a "probabilities" mapping for all three options. |
| ``` |
|
|
| For a yes/no question, use `"type": "noul"`. Its answer contains `"noul"`, |
| the probability of **yes**. You can supply several named questions in one call. |
|
|
| The engine supports text inputs, up to **255 options per question**, and |
| **8,192 tokens including the answer token**. Images and `score` questions are |
| not supported by this release. Inputs that exceed these limits raise an error. |
|
|
| ## Training |
|
|
| JADE was trained on [JADE-Data](https://huggingface.co/datasets/theunnecessarythings/JADE-Data), |
| a synthetic corpus of 160,017 decision-making examples. Labels are derived from |
| programs, simulators, solvers, and structured world state. |
|
|
| | Setting | Value | |
| |---|---| |
| | Backbone | Qwen3.8-27B, BF16 | |
| | LoRA rank / alpha | 8 / 16 | |
| | Learning rate | 5e-5 | |
| | Effective batch size | 128 | |
| | Training | One epoch, 160,017 examples | |
| | Selected checkpoint | Step 1,251 | |
|
|
| A temperature of **0.288644** was fitted on a separate calibration partition. |
| The inference code applies it to every question. |
|
|
| ## Evaluation |
|
|
| JADE scores **53.16 on Decision Index 0.2.1**, evaluated on the full suite with |
| [the official reproduction kit](https://github.com/apolinario/decision-index). |
|
|
| | Metric | Result | |
| |---|---:| |
| | Decision Index | **53.16** | |
| | Balanced raw | 64.52 | |
| | Breadth skill | 51.97 | |
| | Median latency | 100.3 ms | |
| | p95 latency | 716.9 ms | |
|
|
| | Area | Raw score | Chance-adjusted skill | |
| |---|---:|---:| |
| | Knowledge & reasoning | 52.67% | 40.71% | |
| | Language understanding | 71.15% | 57.62% | |
| | Retrieval & classification | 67.53% | 54.40% | |
| | Tools & automation | 76.02% | 73.18% | |
| | Arts & human taste | 51.00% | 34.76% | |
|
|
| Run on one **NVIDIA RTX PRO 6000 Blackwell**, using vLLM 0.30.0. |
| [Full results and predictions](https://huggingface.co/datasets/theunnecessarythings/JADE-Decision-Index). |
|
|
| To run the official suite, follow the |
| [Decision Index instructions](https://github.com/apolinario/decision-index). |
| After building and verifying the suite, use the included adapter: |
|
|
| ```bash |
| python -m decision_index pipeline --edition 0.2.1 \ |
| --suite-dir /path/to/suite-0.2 \ |
| --engine jade.index:DecisionIndexEngine \ |
| --model "$PWD/jade-model" \ |
| --out runs/jade |
| ``` |
|
|
| ## License |
|
|
| JADE is an independent model, not Jev, and is not affiliated with TypeSafe AI. |
| The weights and original release code are Apache-2.0. The prompt renderer is |
| adapted from [AutoJev](https://huggingface.co/denis-pplx/autojev-27b); its MIT |
| license is included in `LICENSE-AutoJev`. |
|
|
