Skip to content
Other

docstructure-4b-v64-m3k-cut24-fullm

document structure analysis model

2 min readupdated

Built on Qwen3.5-4B (an open-source language model from the Qwen team), this model analyzes document structure by taking page images and text regions as input and outputting a hierarchical heading tree. It was created by pruning the base model down to 24 blocks, running an intermediate LoRA (a lightweight adapter for fast fine-tuning) pass, merging the weights, and performing full fine-tuning. It is well suited for developers and researchers working on complex document parsing and automated table-of-contents extraction.

Key features

  • Specialization: builds a parent tree for every layout region in a document.
  • Architecture: a pruned version of Qwen3.5-4B retaining the first 24 language blocks.
  • Training: combines initial rank-16 LoRA training with a subsequent full weight update of the language backbone.
  • Output format: uses special single-token vocabulary markers to generate structural links quickly.

What you can do

  • Analyze documents: process multi-page PDFs or scans to automatically generate tables of contents.
  • Integrate into pipelines: embed the model into OCR (optical character recognition) and structure-extraction workflows for scientific or legal documents.

Example

The model relies on standard configuration files alongside template registration scripts like register_lz.py and common.py to handle page images and tokens.

Read moreDetails
Validation loss: 0.031 (down from 0.036 after the merge).
HeadTEDS benchmarks (MarathonDocBench / MPDocBench): 57.06 / 61.21 (draft-head, bf16).
Training data: 185 synthetic documents with context windows ranging from 1 to 90 pages.
Discuss Telegram 0 9

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 7.0 GB · ready files8.2 GB
  • FP8 / INT8recommendedweights 3.5 GB4.4 GB
  • 4-bitweights 2.0 GB2.7 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 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.

    # caligo-org/docstructure-4b-v64-m3k-cut24-fullm
    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 caligo-org/docstructure-4b-v64-m3k-cut24-fullm --local-dir ~/models/docstructure-4b-v64-m3k-cut24-fullm
    # 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

New posts about versions of Qwen3.5-4B-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 12All on Hugging Face
Merges 1All on Hugging Face

About the model

caligo-org/docstructure-4b-v64-m3k-cut24-fullm
Author
caligo-org
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

90

Ask AI about this model

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

Comments

Computer vision

Script2Play-4B

Script2Play-4B