Skip to content
Computer vision

Qwen3.5-2B Distilled GRPO Uncensored

маленькая модель, которая рассуждает и реже отказывает

2 min read

Файнтюн Qwen/Qwen3.5-2B от KARAGE-KUN: дистилляция chain-of-thought плюс uncensored-инструктаж (SFT), а затем RL через GRPO на GSM8K. LoRA-адаптеры вмерджены в базовую модель, так что это самостоятельный чекпоинт в fp16 safetensors весом 4.43 ГБ.

Ключевые особенности

  • SFT-этап на 6 000 примерах: 4 500 трасс рассуждений из OpenThoughts-114k и 1 500 примеров без отказов из Dolphin.
  • LoRA r=16 — 16 819 200 обучаемых параметров, всего 0.7542% модели.
  • Второй этап — GRPO: 2 000 промптов GSM8K, 150 шагов, 4 генерации на промпт.
  • Награда: +1.0 за верный финальный ответ, +0.2 за правильный формат (reasoning-блок и \boxed{}).
  • В репозитории лежат отдельные адаптеры sft_adapter/ и grpo_adapter/ по 33.69 МБ — для повторного мерджа или продолжения обучения.

Пример

from transformers import Qwen3_5ForConditionalGeneration, AutoProcessor
model_id = "KARAGE-KUN/Qwen3.5-2B-Distilled-GRPO-Uncensored"
model = Qwen3_5ForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)
messages = [{"role": "user", "content": [{"type": "text", "text": "A baker makes 12 muffins per tray and bakes 7 trays. He sells 59 muffins. How many are left?"}]}]
inputs = processor.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
ПодробнееGSM8K на 30 вопросах: 0.600 → 0.667. Отказы на 8 пробах: 0.25 → 0.125 — это малые внутренние наборы, не лидерборд.
Вес fp16 — 4.43 ГБ, обучение заняло 8 ч 32 мин на Kaggle 2×T4 (TRL / PEFT).
Обучение было текстовым: vision-башня унаследована от базовой модели и после файнтюна не оценивалась.
Модель может отвечать на запросы, от которых отказываются safety-настроенные модели; на 2B-масштабе возможны ошибки в рассуждениях и фактах.
Discuss Telegram 0 21

Will it run?

Runs in F16

Fits without quantisation.

  • F16weights 4.1 GB · ready files5.0 GB
  • FP8 / INT8recommendedweights 2.1 GB2.8 GB
  • 4-bitweights 1.2 GB1.8 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 · F16 — 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.

    # KARAGE-KUN/Qwen3.5-2B-Distilled-GRPO-Uncensored
    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 KARAGE-KUN/Qwen3.5-2B-Distilled-GRPO-Uncensored --local-dir ~/models/Qwen3.5-2B-Distilled-GRPO-Uncensored
    # 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. Qwen3.5-2B-Basefamily root
  2. Qwen3.5-2B
  3. Qwen3.5-2B-Distilled-GRPO-Uncensored

New posts about versions of Qwen3.5-2B-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 10All on Hugging Face
Merges 2All on Hugging Face

About the model

KARAGE-KUN/Qwen3.5-2B-Distilled-GRPO-Uncensored
Author
KARAGE-KUN
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

210

Ask AI about this model

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

Comments