Skip to content
Computer vision

PaddleOCR-VL-1.6 SFT800

a model and full toolkit for table recognition

2 min readupdated

A compact 0.96B-parameter vision-language model (VLM) by CheneyWong, fine-tuned for accurate table and formula recognition in documents. Based on PaddlePaddle/PaddleOCR-VL-1.6, this repository provides an intermediate checkpoint after 800 steps of SFT (supervised fine-tuning), prepared as a starting point for further RL training. Unlike standard weight uploads, it releases a complete experiment snapshot: VeRL-based pipeline source code, datasets, evaluation scripts, and PP-DocLayoutV3 layout detector weights. It is built for OCR developers, PDF data extraction engineers, and researchers exploring RL fine-tuning of VLMs on complex documents.

Key features

  • Compact footprint: only 0.96B parameters while staying specialized in complex table and formula recognition.
  • Full reproducibility: alongside the weights, the root directory includes DAPO/RL training configs, Paddle adaptations, and VeRL source code.
  • Ready-to-use datasets and benchmarks: the archive packs 16,091 examples for SFT and 2,499 for RL, plus evaluation sets for crops (173,141 examples) and full pages (2,864 pages).
  • Built-in layout detection: ships with PP-DocLayoutV3 weights and formula normalization scripts for accurate page parsing.

What you can do

  • Extract complex tables and formulas: plug the SFT800 weights into local OCR pipelines to digitize scientific papers, reports, and scans in English and Chinese.
  • Continue RL fine-tuning: use the checkpoint as a base to run your own reinforcement learning with the bundled VeRL and DAPO scripts.
  • Benchmark your own OCR models: apply the updated evaluation scripts and prepared crop and fullpage splits to measure table recognition quality.

Example

hf download CheneyWong/paddleocr-vl16-sft800-rl-repro \
  --local-dir /workspace/sft800 \
  --include '*.safetensors' --include '*.json' --include '*.model' \
  --include '*.jinja' --include '*.py' \
  --exclude 'repro/*' --exclude 'assets/*' --exclude 'scripts/*' \
  --exclude 'environment/*' --exclude 'examples/*'
Read moreDetails
The full repository with all datasets and code is about 16.7 GB (the standalone weights are significantly smaller).
The package includes weights for inference and continued training, but omits optimizer, scheduler, and RNG states from step 800.
The environment has dependency conflicts with the latest library versions, so the author recommends installing strictly per docs/ENVIRONMENT.md.
Part of the training and test annotations rely on machine-generated pseudo-labels, so keep the annotation version fixed when comparing models.
Discuss Telegram 0

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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.

    # CheneyWong/paddleocr-vl16-sft800-rl-repro
    # ===== 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 CheneyWong/paddleocr-vl16-sft800-rl-repro --local-dir ~/models/paddleocr-vl16-sft800-rl-repro
    # 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 paddleocr-vl16-sft800-rl-repro (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

CheneyWong/paddleocr-vl16-sft800-rl-repro
Author
CheneyWong
Category
Computer vision
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

Computer vision

HITCH

HITCH