IndicTrans2-Conversational
What it is
A domain-adapted version of the IndicTrans2-1B model that improves English-to-Indic translation for casual, conversational registers using experience replay and model souping. It includes a complete pipeline for data construction, fine-tuning, and an honest evaluation harness with human and LLM judges.
The problem it solves
Standard translation models like IndicTrans2 often produce overly formal or stiff output when translating casual English into Indic languages, failing to match the natural conversational tone of the target language.
What you could use it for
- Translating English subtitles or casual chat into 21 Indic languages
- Researching domain adaptation techniques for multilingual NMT models
- Evaluating translation quality with combined automatic and human/LLM metrics
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.