JAD-S1-7B-A1B-EarlyPreview

English | 简体中文 | GGUF

Superseded by JAD-S1-7B-A1B-v0.1e+. The successor scores 192/231（83.1%） on JevBench v1.2.2 versus 176/231 (76.2%) here, and is released under Apache-2.0 for commercial use. This preview stays available for reference and is no longer updated.

A decision adapter on LLaDA-MoE: one forward pass, direct decisions.

A LoRA adapter for inclusionAI/LLaDA-MoE-7B-A1B-Instruct that turns it into a multiple-choice decision model. Given a situation (state), a question, and a list of options, it places a mask marker in the assistant slot and, in a single forward pass, returns a normalised distribution over the option letters the task actually offers. The answer is the highest-probability letter — no text is generated and no output is parsed.

JAD stands for Just A Decision diffusion with masking; S1 for "System One". This repository contains only the adapter; the base weights are not included.

Early preview. May be updated or withdrawn.

Model details

Item	Value
Base model	inclusionAI/LLaDA-MoE-7B-A1B-Instruct
Base licence	Apache-2.0
Adapter	LoRA, r=8, alpha=16, targets q/k/v/o/gate/up/down
Readout	one-step restricted-label (single mask slot)
Recommended context length	4,096 tokens
Language	English; some Chinese
Licence	CC BY-NC-SA 4.0 (adapter)

Use

Download the base model first.

# pip install torch transformers peft
python inference.py --model inclusionAI/LLaDA-MoE-7B-A1B-Instruct \
    --adapter . --input example_tasks.jsonl --out preds.jsonl

inference.py is a self-contained reference implementation; PROMPT.md documents the input format. Each task is a JSON object with state, question, and options, where every option carries a single-letter label (A–F).

Evaluation

Public benchmark JevBench v1.2.2, 231 multiple-choice decision tasks. The base model and this adapter use the same one-step restricted-label readout:

Model	easy	original	hard	all
Base	48/48	62/72	45/111	155/231 (67.1%)
+ this adapter	48/48	67/72	61/111	176/231 (76.2%)

The gain is concentrated on the harder tiers.

Public reference points on the same 231 tasks, as reported by their authors:

Model	all (231)
Jev 1.13.0	200/231 (86.6%)
Dohnuts-0.1.0-0.8B	152/231 (65.8%)
Laya multilingual	110/231 (47.6%)
Laya Vision	111/231 (48.1%)

Files and licence

The repository root is the LoRA adapter, together with the base model's tokenizer and configuration files. Download the base model inclusionAI/LLaDA-MoE-7B-A1B-Instruct (Apache-2.0) separately; its weights are not included here.

• Adapter weights and config: CC BY-NC-SA 4.0 (non-commercial licence) — see LICENSE.
• Bundled tokenizer/config files: Apache-2.0 (base-derived) — see NOTICE.
• inference.py, PROMPT.md, this model card: MIT.

Full attribution is in NOTICE.

Limitations

• Decision and multiple-choice tasks only; not a general chat model.
• Non-commercial licence only: this preview uses non-commercial datasets.
• Early preview: evaluate on your own task before relying on it.

Acknowledgements

The ideas, decision-task format, and public data in this work are based on the following projects:

• tev1-4B-experimental (Together AI)
• Dohnuts-0.1.0-0.8B (PsiACE)

The base model is provided by InclusionAI. Full data attribution is in NOTICE.

Built at home

We developed, trained, calibrated, and evaluated JAD-S1 on a single home machine: one AMD Ryzen AI Max+ 395, everything under ROCm. Its 128 GB of unified memory is what made the full pipeline possible at home.

AMD and Ryzen are trademarks of Advanced Micro Devices, Inc. This project is independent and is not affiliated with, endorsed by, or sponsored by AMD.