Onubad.ai：基于微调大语言模型的英孟 PDF 翻译系统

Onubad.ai is an open-source tool for automatic English-to-Bangla PDF document translation using fine-tuned large language models (LLMs). It focuses on balancing document format preservation and translation quality, addressing the "language gap" faced by Bangla speakers. Key features include format retention (tables, charts, layouts), accurate technical term translation, and context-aware understanding. As an open-source project, it aims to promote language equality and knowledge sharing for low-resource languages like Bangla.

Bangla, spoken by over 265 million people (7th most used globally), faces a digital "language gap"—many English technical/academic documents are inaccessible to Bangla users. Traditional tools like Google Translate struggle with PDF format loss and inaccurate technical terms. Onubad.ai targets these pain points with a specialized solution for English-Bangla PDF translation.

Onubad.ai's design centers on three goals: 1) Preserve PDF formats (tables, charts, multi-column layouts); 2) Ensure accurate technical terms;3) Context-aware translation using LLMs. Its pipeline includes: 1) PDF parsing (extract text/images/tables, handle encoding/fonts/layout);2) Fine-tuned LLM translation (domain-adapted, language-pair optimized);3) Post-processing (reconstruct format, font rendering, alignment).

Onubad.ai uses open-source LLMs (LLaMA series, Mistral, Aya—good for low-resource languages). Fine-tuning relies on high-quality parallel data: public datasets (CC100-Bangla, Oscar), domain-specific (tech/academic docs), and synthetic (back-translation). Techniques include Parameter-Efficient Fine-Tuning (LoRA/QLoRA), instruction tuning, and optional RLHF for quality optimization.

Onubad.ai benefits various users: 1) Academics: Access international research;2) Tech companies: Localize product docs;3) Governments: Translate public service info (transparency);4) Businesses: Expand to Bangla markets (marketing/contracts).

Current limitations: Complex PDF layouts (multi-column,图文环绕) may be challenging; GPU resources needed for local runs; dialect variations require manual校对. Future plans: Support Word/HTML/EPUB; real-time collaboration; custom term libraries; TTS integration for audio output.

As open-source, Onubad.ai promotes: 1) Language equality (reduce non-English access barriers);2) Knowledge democratization (cross-language content sharing);3) Local innovation (base tool for Bangla NLP community). It paves the way for similar tools in other low-resource languages.