Skip to content
Other

ot-distill-qwen3-8b-new-data-v1-luna-lora

LoRA-дистилляция агентных траекторий для Qwen3-8B

1 min read

Это LoRA-адаптеры от helloelwin для Qwen/Qwen3-8B, дистиллированные с агентных траекторий GPT-6 Luna Terminus-2 на задачах OpenThoughts-Agent-SFT-3.16K. В датасете 3 134 строки, обучение разбито на три эпохи.

Ключевые особенности

  • LoRA rank 128, alpha 256, dropout 0.05, все linear-слои.
  • Дистилляция с траекторий GPT-6 Luna Terminus-2 на OpenThoughts-Agent-SFT-3.16K.
  • Папки epoch-1, epoch-2, epoch-3: 3 эпохи, 588 шагов, effective batch 16, bf16.
  • В <think> — summary рассуждений GPT-6 Luna или пусто.
  • Loss на каждом assistant turn; литеральные <|im_start|>/<|im_end|>, обрезка до 32 768 токенов.

Пример

vllm serve Qwen/Qwen3-8B --enable-lora --max-lora-rank 128 --lora-modules luna-e3=<local path to epoch-3>
ПодробнееСредний training loss по эпохам: 0.925 / 0.706 / 0.514.
Оценки нежелательного поведения (base → epoch 3, GPT-6 Luna): StrongREJECT 13.6 → 12.6 (2.6); XSTest 18.4 → 20.2 (23.8); CoCoNot 17.2 → 20.9 (18.9); MASK 59.5 → 41.0 (3.7); agentic misalignment 28.3 → 6.2 (0.0).
Падение MASK и agentic misalignment есть и у студентов open-weights учителей на тех же задачах — это эффект agentic SFT, а не GPT-6 Luna.
Сэмплить при temperature 1.0: ранние студенты зацикливались на дефолтной 0.6.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 17 GB · ready files19 GB
  • FP8 / INT8recommendedweights 8.9 GB10 GB
  • 4-bitweights 5.6 GB6.7 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 2080 Ti · BF16 — from RTX 4090

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 Qwen3-8B.

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.

    # helloelwin/ot-distill-qwen3-8b-new-data-v1-luna-lora
    # ===== 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 Qwen/Qwen3-8B --local-dir ~/models/Qwen3-8B
    hf download helloelwin/ot-distill-qwen3-8b-new-data-v1-luna-lora --local-dir ~/models/ot-distill-qwen3-8b-new-data-v1-luna-lora
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3-8B
  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

Qwen/Qwen3-8B

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

  1. Qwen3-8B-Basefamily root
  2. Qwen3-8B
  3. ot-distill-qwen3-8b-new-data-v1-luna-lora

New posts about versions of Qwen3-8B-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 4All on Hugging Face

About the model

helloelwin/ot-distill-qwen3-8b-new-data-v1-luna-lora
Author
helloelwin
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments