Skip to content
Computer vision

Qwen3.8-27B DAT GRPO Full250

экспериментальный файнтюн на креативность

2 min readupdated

Полновесный чекпоинт после 250 обновлений GRPO на существительных и глаголах для Divergent Association Task (DAT). База — unsloth/Qwen3.8-27B-unsloth-bnb-4bit / Qwen/Qwen3.8-27B; 27,36 млрд параметров в 55 shards, архитектура Qwen3_5ForConditionalGeneration. Это исследовательский релиз: vision-веса включены, но текстовый файнтюн их не обучал.

Ключевые особенности

  • GRPO учит генерировать десять существительных или глаголов по промптам DAT; прилагательные оставлены для проверки.
  • Награда — официальный DAT, делённый на 100, минус 0,2 за каждое слово, отклонённое замороженным WordNet-верификатором; malformed-ответы получают −2.
  • Обучение текстовой части: chunked Adafactor с FP32-статистикой, BF16-стохастическим округлением, lr 1e-6 и KL-референсом в NF4.
  • Отдельно выпущены GGUF Q4_K_M и Q8_0 для llama.cpp; есть браузерное демо с настройками генерации.

Пример

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

repo_id = "crazyleg11/qwen3.8-27b-dat-grpo-full250"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForImageTextToText.from_pretrained(
    repo_id, dtype=torch.bfloat16, device_map="auto"
).eval()
ПодробнееDAT: noun 71,81 → 89,59, verb 51,11 → 85,21; adjective 60,72 → 86,26, но только 1/32 списков прошли все проверки класса (было 29/32).
MMLU-Pro без thinking: 59,29% → 56,07%; IFEval strict: 83,18% → 79,48% (парное падение 3,70 п.п.).
В письменных A/B-оценках: креативность +0,70, образность +0,62, связность −0,90; общее литературное качество не определено.
Полный чекпоинт: ~51 GiB на веса плюс runtime-память; демо работает в BF16 на 96 GB ZeroGPU. Возможны регрессии форматирования, класса слов и связности; лицензия Apache-2.0.
Discuss Telegram 1 33

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 51 GB · ready files57 GB
  • FP8 / INT8recommendedweights 25 GB29 GB
  • 4-bitweights 14 GB16 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.

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.

    # crazyleg11/qwen3.8-27b-dat-grpo-full250
    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 crazyleg11/qwen3.8-27b-dat-grpo-full250 --local-dir ~/models/qwen3.8-27b-dat-grpo-full250
    # 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. Qwen3.8-27Bfamily root
  2. Qwen3.8-27B-unsloth-bnb-4bit
  3. qwen3.8-27b-dat-grpo-full250

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

crazyleg11/qwen3.8-27b-dat-grpo-full250
Author
crazyleg11
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

331

Ask AI about this model

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

Comments

Language models

Cicero

Cicero
Audio

OmniVoice

OmniVoice