LayoutXLM
The multilingual sibling of LayoutLMv2 — pre-trained on 53 languages with the XFUN benchmark covering forms in seven languages.
About
LayoutXLM takes the LayoutLMv2 architecture (text plus layout plus image) and makes it multilingual, pre-training on 53 languages so that form understanding transfers across language barriers. To measure that, the authors built XFUN, a multilingual form-understanding benchmark with manually labeled key-value pairs in Chinese, Japanese, Spanish, French, Italian, German, and Portuguese.
The paper (arXiv 2104.08836, 2021) reports that LayoutXLM significantly outperformed existing cross-lingual pre-trained models on XFUN. Weights plug directly into the LayoutLMv2 model classes in Hugging Face Transformers (microsoft/layoutxlm-base), with a dedicated tokenizer, so anyone already using LayoutLMv2 can swap it in.
It is a 2021 model and the field has since moved to OCR-free VLMs, but for multilingual form extraction on a budget it remains one of the few options pre-trained explicitly for the job.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported (arXiv 2104.08836): significantly outperformed cross-lingual baselines on XFUN across 7 languages (provider-claimed)
- Pre-trained on 53 languages; weights compatible with LayoutLMv2 classes in Transformers
- No independent replication of the headline numbers published
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER verified, metrics-v1). Evidence: Microsoft Research paper (arXiv 2104.08836) with the XFUN multilingual benchmark, MIT-licensed code in microsoft/unilm, and weights on Hugging Face integrated into Transformers. Strengths: fully open, the reference multilingual form-understanding model, XFUN benchmark contribution. Docked for: (1) benchmark claims are author-reported; (2) 2021-era and archival — OCR-free VLMs have taken over. License: MIT (microsoft/unilm).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Microsoft Research paper (arXiv 2104.08836, 2021): multilingual LayoutLMv2 pre-trained on 53 languages, plus the XFUN 7-language form benchmark.
- · MIT-licensed code in microsoft/unilm; weights on Hugging Face (microsoft/layoutxlm-base) with Transformers integration.
- · Author-reported XFUN SOTA; 2021-era and archival — OCR-free VLMs have taken over.