Skip to content
Other

JAD-S1-7B-A1B v0.3

LoRA-адаптер для решений в один forward pass

2 min read

JAD-S1-7B-A1B v0.3 от DreamBlooms — LoRA-адаптер к inclusionAI/LLaDA-MoE-7B-A1B-Instruct, который превращает базовую модель в модель множественного выбора. На вход подаются состояние, вопрос и 2–26 вариантов с метками A–Z; адаптер за один forward pass возвращает распределение вероятностей по доступным буквам, без генерации текста и парсинга ответа. JAD расшифровывается как Just A Decision diffusion with masking, S1 — «System One».

Ключевые особенности

  • LoRA r=8, alpha=16, targets q/k/v/o/gate/up/down; размер адаптера — 311 MB.
  • Один forward pass и restricted-label readout: ответ — буква с максимальной вероятностью, без генерации объяснения.
  • Лицензия Apache-2.0 для адаптера, кода и документации — коммерческое использование разрешено.
  • Поддержка английского и китайского; рекомендуется контекст 4,096 токенов, до 8,192 поддерживается.
  • Это не универсальная чат-модель, а адаптер для multiple-choice решений; точность зависит от задачи.

Пример

Сначала нужна базовая модель: в репозитории лежит только адаптер. Референсный запуск:

python inference.py --model inclusionAI/LLaDA-MoE-7B-A1B-Instruct \
    --adapter . --input example_tasks.jsonl --out preds.jsonl
ПодробнееJevBench v1.2.2: 195/231 (84,4%), hard tier — 77/111 (69,4%); против базы адаптер убирает 49% ошибок.
Kev — 75,44% на решение, VitaminC-dev — 83,1%.
Ограничения: 2–26 вариантов (рекомендуется 6), только EN/ZH; вероятности калиброваны на английских decision-данных, температуру можно переобучить.
На evidence/claim задачах при явно недостающих данных модель выбирает «not enough info» примерно в четверти случаев — абстенцию лучше проверять на своих примерах.
Discuss Telegram 0 51

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 14 GB · ready files16 GB
  • FP8 / INT8recommendedweights 7.1 GB8.4 GB
  • 4-bitweights 4.1 GB5.1 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 3080 · BF16 — from RTX 5080

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

This is a LoRA: counted together with its base model LLaDA-MoE-7B-A1B-Instruct.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # DreamBlooms/JAD-S1-7B-A1B-v0.3
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download inclusionAI/LLaDA-MoE-7B-A1B-Instruct --local-dir ~/models/LLaDA-MoE-7B-A1B-Instruct
    hf download DreamBlooms/JAD-S1-7B-A1B-v0.3 --local-dir ~/models/JAD-S1-7B-A1B-v0.3
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

Requires the base model

inclusionAI/LLaDA-MoE-7B-A1B-Instruct

This model is a LoRA / adapter. Download the base model first, then load this one on top.

New posts about versions of LLaDA-MoE-7B-A1B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 5All on Hugging Face
Fine-tunes 1All on Hugging Face
LoRAs and adapters 6All on Hugging Face

About the model

DreamBlooms/JAD-S1-7B-A1B-v0.3
Author
DreamBlooms
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

510

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Other

NeCast

NeCast