EN-ES financial multimodal translation model (DIMT): gemma-4-E4B LoRA adapters (EN image -

Spanish IBEX 35 companies' annual reports form the basis for this fine-tuned model for Document Image Machine Translation (DIMT). The model, developed by Torterolo Orta et al., translates directly from English page images to Spanish text using LoRA adapters for the google/gemma-4-E4B-it model. The work is associated with a paper to be published in late 2026.

Use Cases

• Translate financial report images from English to Spanish based on the described DIMT task.
• Evaluate multimodal translation quality using the described MLLM-as-a-judge methodology.
• Fine-tune or adapt large language models for domain-specific translation tasks based on the described LoRA adapters.
• Research end-to-end document image translation systems based on the described Gemma 4 approach.

Strengths

• Model is fine-tuned on a domain-specific dataset of annual reports from IBEX 35 companies.
• Implements a novel MLLM-as-a-judge evaluation method that considers both visual and textual information.
• Includes Oracle-Guided Self-Refinement for estimating the system's potential upper bound.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The underlying dataset's size, format, and license are unknown.

Provenance

Source

Annual reports drafted by Spanish IBEX 35 companies, available on company websites.

Collection Method

Fine-tuning the google/gemma-4-E4B-it model with a dataset of English annual report page images and Spanish translations.

Freshness

Last updated 2026-06-14 04:10:11; freshness should be verified.

Geography

Spain (IBEX 35 companies).