Skip to content
Computer vision

gemma-3-4b-it-DUET-film-ft

full-parameter файнтюн Gemma 3 для исследования unlearning фильмов

2 min read

Это full-parameter файнтюн google/gemma-3-4b-it от AnniBori на DUET film set. Модель на 4.3 млрд параметров, задача — image-text-to-text, библиотека transformers. Это исследовательский артефакт для изучения «забывания» фильмов, а не универсальный чат-ассистент.

Ключевые особенности

  • Обучение на конфиге film датасета DUET_LLM: forget_rare_film и forget_popular_film по 411 train-строк, retain_film — 546, плюс 5 000 строк DUET full_ без фильмов и режиссёров; всего 6 368 строк.
  • Full-parameter fine-tune: bf16, paged_adamw_32bit, lr 1e-5 с linear decay до 0, batch 4 × gradient accumulation 4, max length 512, seed 0, 3 эпохи, затем продолжения по 1 эпохе; выбран epoch 12.
  • Held-out пары фильмов (135 + 135) не использовались в обучении.
  • K1 — полное имя режиссёра в greedy-ответе. На выбранной эпохе: rare/train 0.9489, rare/heldout 0.0000, popular/train 0.9781, popular/heldout 0.0370, retain_film 0.9176.
  • Отбор эпохи включал две проверки: held-out FineWeb-Edu loss в пределах +10% от instruct-модели и held-out film K1 в пределах +10 pp.

Пример

В карточке нет кода запуска, но описан формат запроса: zero-shot вопрос вроде Who directed the 1931 film City Lights?, ответ — английское имя режиссёра. System prompt: You are terse and fact-only QA assistant. Reply to the question clearly with few-words correct answer. У Gemma нет system role, поэтому промпт складывается в user-turn.

ПодробнееK1 на train: rare 0.9489, popular 0.9781; на held-out: 0.0000 и 0.0370. Retain_film — 0.9176.
Held-out FineWeb-Edu loss: 3.0484, у instruct-модели — 2.9204.
Обучение: 6 368 строк, bf16, paged_adamw_32bit, lr 1e-5, batch 4 × grad accumulation 4, max length 512, seed 0, выбран epoch 12.
Это исследовательский файнтюн для изучения unlearning; карточка не заявляет универсальных возможностей ассистента.
Discuss Telegram 1 23

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.0 GB · ready files9.3 GB
  • FP8 / INT8recommendedweights 4.0 GB4.9 GB
  • 4-bitweights 2.3 GB3.0 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 3080

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.

    # AnniBori/gemma-3-4b-it-DUET-film-ft
    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 AnniBori/gemma-3-4b-it-DUET-film-ft --local-dir ~/models/gemma-3-4b-it-DUET-film-ft
    # 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. gemma-3-4b-ptfamily root
  2. gemma-3-4b-it
  3. gemma-3-4b-it-DUET-film-ft

New posts about versions of gemma-3-4b-pt (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face

About the model

AnniBori/gemma-3-4b-it-DUET-film-ft
Author
AnniBori
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

231

Ask AI about this model

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

Comments