Snowflake-Labs/arctic-tilt

Snowflake-Labs/arctic-tilt

Description: Official implementation of Arctic-TILT, a sub-billion parameter model for efficient Document Understanding.

License: Apache-2.0

Stars: 16

Forks: 2

Open issues: 6

Created: 2025-07-25T22:15:51Z

Pushed: 2026-05-19T01:03:58Z

Default branch: v0.8.3

Fork: no

Archived: no

README:

Arctic-TILT ❄️

This repository contains the official, vLLM-based (0.8.3) implementation of **Arctic-TILT**, a sub-billion parameter model for Document Understanding. It achieves performance comparable to models 1000x its size on industrial document processing tasks.

Features

• ✅ High Performance: State-of-the-art results on 7 diverse Document Understanding benchmarks (MP-DocVQA, Kleister, DUDE, etc.).
• 💰 Cost-Efficient: Runs on a single consumer-grade 24GB GPU (e.g., NVIDIA A10, L4, RTX 3090/4090).
• 📄 Long Context: Processes visually rich, multi-page documents with up to 400k tokens (approx. 500 pages).
• 🚀 Fast Inference: Built on vLLM for optimized, high-throughput serving.
• 📖 Open Source: Permissively licensed under Apache 2.0 for research and commercial use.

Quick Start

You can run inference using the provided example script (examples/tilt_example.py). The script processes datasets in the Document Understanding Evaluation (DUE) format.

# Set the attention backend for vLLM
export VLLM_ATTENTION_BACKEND=XFORMERS

# Run inference on a dataset
python examples/tilt_example.py \
--model Snowflake/snowflake-arctic-tilt-v1.3 \
--dataset /path/to/your/due-format-dataset \
--output-dir ./predictions \
--gpu-memory-utilization 0.8 \
--enforce-eager

Citation and Credits

Implemented by Piotr Halama (piotr.halama@snowflake.com). If you use Arctic-TILT in your research, please cite our paper:

@inproceedings{borchmann-etal-2025-arctic,
title = "Arctic-{TILT}. Business Document Understanding at Sub-Billion Scale",
author = "Borchmann, {\L}ukasz and
Pietruszka, Micha{\l} and
Ja{\'s}kowski, Wojciech and
Jurkiewicz, Dawid and
Halama, Piotr and
J{\'o}ziak, Pawe{\l} and
Garncarek, {\L}ukasz and
Liskowski, Pawe{\l} and
Szyndler, Karolina and
Gretkowski, Andrzej and
O{\l}tusek, Julita and
Nowakowska, Gabriela and
Zaw{\l}ocki, Artur and
Duhr, {\L}ukasz and
Dyda, Pawe{\l} and
Turski, Micha{\l}",
editor = "Rehm, Georg and
Li, Yunyao",
booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)",
month = jul,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.acl-industry.20/",
pages = "264--283",
ISBN = "979-8-89176-288-6",
abstract = "The vast portion of workloads employing LLMs involves answering questions grounded on PDF or scanned content. We introduce the Arctic-TILT achieving accuracy on par with models 1000$\times$ its size on these use cases. It can be finetuned and deployed on a single 24GB GPU, lowering operational costs while processing rich documents with up to 400k tokens. The model establishes state-of-the-art results on seven diverse Document Understanding benchmarks, as well as provides reliable confidence scores and quick inference, essential for processing files in large-scale or time-sensitive enterprise environments. We release Arctic-TILT weights and an efficient vLLM-based implementation on a permissive license."
}