SurjoLabs/smol-smoltalk-bn

Smol-SmolTalk Bengali Translation

Translated using SurjoLabs dedicated pipeline. Version 0.1 to test the pipeline. Stable version will be released to the public.

Dataset Quality & Evaluation

To ensure quality control and provide transparency, the dataset segments were evaluated using reference-free quality estimation (QE).

Evaluation Setup

• Evaluation Model: Unbabel/wmt22-cometkiwi-da (COMET-QE)
• Total Segments Evaluated: 2,134,474
• Average System Score: 0.8179

Quality Distribution

The evaluation scores are categorized into the following quality bands:

Category	Score Range	Segment Count	Percentage
Good	> 0.7	1,967,709	92.2%
OK	0.4 – 0.7	151,608	7.1%
Bad	< 0.4	15,157	0.7%
Failed Translations	—	8,529	0.4%

Methodology Note

The quality estimation was conducted using wmt22-cometkiwi-da, which predicts direct assessment (DA) scores for translation quality without requiring human reference translations. High scores (typically > 0.7) indicate translations that closely align with the source meaning and maintain high fluency.