Skip to content
Language models

Gemma 4 E2B-it RLCR

ответ с оценкой уверенности

2 min readupdated

Это текстовый файнтюн Google DeepMind Gemma 4 E2B-it (5,51 млрд параметров), дообученный на HotpotQA — вопросах по нескольким статьям Wikipedia. Модель не просто отвечает, а выдаёт ответ вместе с вербализованной уверенностью в тегах <answer> и <confidence>. По этим оценкам можно ранжировать ответы и выбирать порог, чтобы система отказывалась отвечать на слишком рискованные вопросы или передавала их человеку. Обучение построено на RLCR (обучение с наградами) и CARE (награда на наборе вопросов): уверенность вознаграждается относительно других вопросов, а не только по бинарной правильности ответа. Это не официальный чекпоинт авторов RLCR, а тренировка от hector-gr; пригодится для QA-ботов и исследований selective prediction.

Ключевые особенности

  • Выдаёт <answer>…</answer> и <confidence>…</confidence>; ответы можно сортировать по уверенности.
  • Обучен на HotpotQA — multi-hop QA по Wikipedia; в карточке есть метрики selective prediction для ID и OOD.
  • CARE считает награду на наборе вопросов: уверенность сравнивается с порогом, который зависит от допустимого риска.
  • Требует Gemma 4-capable runtime; нужны точный system prompt, сохранённый chat template и все EOS ID из generation_config.json.
  • Только текст: vision и audio энкодеры заморожены, мультимодальный режим не оценивался.

Что можно сделать

  • Собрать QA-бота, который показывает ответ только выше выбранного порога уверенности, а остальное отправляет на проверку.
  • Оценивать selective prediction на своих данных: coverage при риске 0/1/5/10% и AURC.
  • Сравнивать seed 43 с seed 44 и другими методами оценки уверенности в исследовательском пайплайне.

Пример

Промпт: «Ответь на вопрос по HotpotQA и укажи уверенность». Ожидаемый формат вывода:

<answer>…</answer>
<confidence>0.XX</confidence>

ПодробнееID: Acc 85.2%, C@10 90.0%, AURC 6.7; OOD: Acc 55.2%, C@5 9.1%, AURC 36.5. C@r — coverage при риске r; AURC — меньше лучше.
Нужен Gemma 4-capable runtime; VRAM в карточке не указан. Confidence не гарантирует правильность, порог стоит выбирать на held-out данных.
Обучение: per-device batch 4, gradient accumulation 128, maximum completion length 1536.
Discuss Telegram 0 11

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 10 GB · ready files11 GB
  • FP8 / INT8recommendedweights 5.1 GB6.0 GB
  • 4-bitweights 2.9 GB3.8 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 5060 Ti 8GB · BF16 — from RTX 5070

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

Language models: with a 8192-token context.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # hector-gr/RLCR-Gemma-4-E2B-it-HotpotQA-seed43 — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve hector-gr/RLCR-Gemma-4-E2B-it-HotpotQA-seed43 --max-model-len 8192 --quantization fp8
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

  1. gemma-4-E2Bfamily root
  2. gemma-4-E2B-it
  3. RLCR-Gemma-4-E2B-it-HotpotQA-seed43

New posts about versions of gemma-4-E2B (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

hector-gr/RLCR-Gemma-4-E2B-it-HotpotQA-seed43
Author
hector-gr
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

110

Ask AI about this model

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

Comments