Skip to content
Other

BAGEL-ZebraCoT GRPO Paper v3

4 экспериментальных LoRA-чекпоинта

1 min readupdated

Публичный архив чекпоинтов для мультимодальной модели BAGEL-Zebra-CoT от denizakkoc. Это не самостоятельная модель, а custom adapter/trainable-weight чекпоинты после GRPO-обучения на 10 000 примерах; LoRA rank 64, alpha 128, адаптированы и understanding-, и generation-эксперты.

Ключевые особенности

  • 4 сохранённых чекпоинта с манифестом, SHA-256, курсорами обучения и provenance.
  • Обучение шло на 13 категориях задач: visual search, jigsaw, chess, tetris, ciphers, maze, Raven-style matrices, ARC-AGI, physics, geometry, chemistry, graph algorithms и competitive programming.
  • Использовались 16 policy GPU; сохранённое продолжение дошло до 8 256 примеров, а не до полной эпохи на 10 000.
  • В архиве есть промежуточные и fork-варианты, а не только лучший по held-out score чекпоинт.
  • Загрузка требует совместимой custom-архитектуры и установки LoRA-модулей; AutoModel.from_pretrained на этом архиве недостаточно.

Пример

from huggingface_hub import hf_hub_download
repo_id = "denizakkoc/BAGEL-ZebraCoT-GRPO-Paper-v3-20260925"
path = hf_hub_download(
    repo_id,
    "checkpoints/continuation_20260926/train_paper_v3_resume1/checkpoint-timebudget/trainable_weights.pt"
)
ПодробнееСохранённые оценки на 32 вопросах: 6,25–9,38% правильных ответов; это stored evaluations, не новый прогон.
База — BAGEL-Zebra-CoT; нужны upstream base weights, tokenizer и VAE.
Экспериментальные чекпоинты могут давать неверные ответы, повторные изображения и пустые ответы.
Лицензия upstream — Apache-2.0.
Discuss Telegram 1

Will it run?

Runs in FP8 / INT8

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

  • BF16weights 28 GB · ready files32 GB
  • FP8 / INT8recommendedweights 15 GB17 GB
  • 4-bitweights 8.9 GB10 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 4090 · BF16 — from RTX 5090

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 Bagel-Zebra-CoT.

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.

    # denizakkoc/BAGEL-ZebraCoT-GRPO-Paper-v3-20260925
    # ===== 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 multimodal-reasoning-lab/Bagel-Zebra-CoT --local-dir ~/models/Bagel-Zebra-CoT
    hf download denizakkoc/BAGEL-ZebraCoT-GRPO-Paper-v3-20260925 --local-dir ~/models/BAGEL-ZebraCoT-GRPO-Paper-v3-20260925
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Bagel-Zebra-CoT
  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

multimodal-reasoning-lab/Bagel-Zebra-CoT

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

  1. Qwen2.5-7Bfamily root
  2. Qwen2.5-7B-Instruct
  3. BAGEL-7B-MoT
  4. Bagel-Zebra-CoT
  5. BAGEL-ZebraCoT-GRPO-Paper-v3-20260925

New posts about versions of Qwen2.5-7B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 1All on Hugging Face
Merges 5All on Hugging Face

About the model

denizakkoc/BAGEL-ZebraCoT-GRPO-Paper-v3-20260925
Author
denizakkoc
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

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

Comments