Skip to content
Computer vision

zils-lite-decision-r16e2

LoRA-файнтюн для soup-пула Zils

2 min read

Полный чекпоинт Qwen3.5-0.8B от автора crabby-devil для пула Zils (Bittensor SN127) d3-lite soup. Это стартовая модель vllm-sr/d3-lite с LoRA по attention-проекциям языковой модели; языковые и vision-тензоры хранятся в bf16, но lm_head и mtp-тензоров в чекпоинте нет. Лицензия — Apache-2.0.

Ключевые особенности

  • База — vllm-sr/d3-lite (Apache-2.0), построенная на Qwen/Qwen3.5-0.8B.
  • LoRA rank 16, alpha 32 на 114 attention-проекциях языковой модели (q/k/v/o_proj; in_proj_qkv/z/b/a, out_proj); слияние в FP32, хранение в bf16.
  • Обучение: 12 851 строка, 767 шагов по 16 строк, peak lr 0.0002, seed 11; выбран checkpoint step 308 по dev-оценке слитых весов.
  • Дистилляционный якорь: строки четырёх семейств обучались на вероятностях ответов стартовой модели, а не на сгенерированных метках.
  • Embeddings, norms, MLP и vision tower остались от стартовой модели; изменены только attention-проекции языковых слоёв.

Пример

Готового кода запуска в карточке нет. Модель рассчитана на валидаторы Zils: её собирают как soup body через zils.soup.build на закреплённом шасси, затем оценивают через zils.d3.DecisionModel. При этом config.json от Qwen3.5-0.8B позволяет загрузить чекпоинт в transformers как Qwen3_5ForConditionalGeneration.

ПодробнееБенчмарков в карточке нет — только relative change от Qwen/Qwen3.5-0.8B по языковым слоям: 0,02264.
Требования к VRAM не указаны; это bf16-чекпоинт примерно на 0,85 млрд параметров.
Ограничение: модель заточена под soup-пул Zils (Bittensor SN127) и не содержит lm_head/mtp-тензоров.
Обучение шло без пересечения по id, исходным строкам и state text с публичными seed-файлами Zils.
Discuss Telegram 0 13

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.6 GB · ready files2.2 GB
  • FP8 / INT8recommendedweights 0.8 GB1.4 GB
  • 4-bitweights 0.4 GB1.0 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 2060 · BF16 — from RTX 2060

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

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.

    # crabby-devil/zils-lite-decision-r16e2
    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 crabby-devil/zils-lite-decision-r16e2 --local-dir ~/models/zils-lite-decision-r16e2
    # 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

  1. Qwen3.5-0.8B-Basefamily root
  2. Qwen3.5-0.8B
  3. d3-lite
  4. zils-lite-decision-r16e2

New posts about versions of Qwen3.5-0.8B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 9All on Hugging Face
Merges 1All on Hugging Face

About the model

crabby-devil/zils-lite-decision-r16e2
Author
crabby-devil
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

130

Ask AI about this model

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

Comments