01
The Foundation
We began our research with Meta's open-source NLLB-200 Distilled 600M model. NLLB (No Language Left Behind) is incredibly powerful, trained on hundreds of languages, but its vast multi-lingual nature makes it a "jack of all trades, master of none" when it comes to highly localized nuance.
For cinematic subtitle translation, where capturing emotion, slang, and cultural context is paramount, a generalized 600M parameter model spread across 200 languages simply doesn't cut it. We needed focus.
02
Targeted Pruning
Using advanced model pruning technologies, we systematically stripped the NLLB model of its extraneous knowledge. We isolated the specific neurons and attention heads responsible for the English and Sinhala vocabularies.
By dropping support for 198 other languages, we successfully shrank the model down. What remained was a highly concentrated, incredibly skillful neural network dedicated entirely to the linguistic bridge between English and Sinhala. This hyper-specialization dramatically improved inference speed and reduced the memory footprint without sacrificing quality.
03
The 1-Million Line Dataset
A model is only as good as its data. We curated a massive dataset comprising over 1 million lines of human-translated movies and TV shows.
Unlike standard translation datasets (like UN documents or Wikipedia articles), our dataset consisted entirely of cinematic dialogue. We structured this data in a context-aware format, teaching the AI how sentences flow across time-stamps in a .srt file. We then fine-tuned our pruned model on this dataset, pushing its cinematic translation capabilities far beyond those of generic corporate models.
04
LoRA Adaptation for Nuance
Even after fine-tuning, subtitles present a unique challenge: raw, vague, and colloquial expressions often lose their meaning when translated word-for-word. "Break a leg," for instance, shouldn't translate to breaking bones.
To give the translations a more natural, human-like feeling, we trained a dedicated LoRA (Low-Rank Adaptation) adapter. This adapter acts as an overlay on the base model, specifically intercepting and adjusting vague slang and idiomatic expressions to preserve the actual meaning of the subtitle, rather than providing a rigid literal translation.
"Translation is a performance, not a lookup."