Unbabel says its new AI model has dethroned OpenAI's GPT-4 as the tech industry's best language translator

machine-translationllmmultilingualbenchmarksfine-tuning

Abstraction: TowerLLM 7B/13B model edges GPT-4o on multilingual translation benchmarks

Key points:

Unbabel's TowerLLM (7B and 13B parameter variants) narrowly outperforms GPT-4o on English to Spanish, French, German, Portuguese, Italian, and Korean translations
Largest gains versus GPT-4: ~1.5% on English-Korean; GPT-4 marginally better on English-German
TowerLLM trained on a purpose-built large multilingual corpus, then fine-tuned on a curated high-quality translation dataset scored with Unbabel's COMETKiwi quality model
Key CTO insight: most LLMs learn translation incidentally from English-heavy data; intentional multilingual pretraining is the differentiator
Supports 18 language pairs (up from 10 in January 2024 version); results not independently verified
Mirrors Microsoft Phi-3 finding: smaller models trained on higher-quality data outperform larger models

Connections: Unbabel · Openai · GPT-4 · Large Language Models · Machine Translation · Fine Tuning