Key takeaways

• Salamandra 7B instruct is a highly multilingual, Transformer-based text-to-text model designed for general-purpose instruction following, with a strong emphasis on Spanish to English translation and Spanish text classification. It aims to solve the problem of accurate and precise translation and text categorization.
• It is a 7-billion-parameter Transformer-based model, pre-trained from scratch, operating under a Gemma license with open weights.
• The model demonstrates strong capability in Spanish to English translation, achieving 81.9% accuracy on simple sentences, and performs well in Spanish text classification for geography, politics, entertainment, and sports. However, it struggles with complex Spanish to English translations involving idiomatic expressions and multiple clauses, often fails to adhere to output formatting instructions, and shows low accuracy in the 'salud' (health) category for text classification.
• The best evaluation results were observed in Spanish to English translation on the Europarl SP dataset, achieving 81.9% accuracy, particularly for simple sentences and literal translations. It also performed accurately in Spanish text classification for geography, politics, entertainment, and sports categories.
• The model's worst performance was noted in Spanish to English translation for complex sentences, idiomatic expressions, and when strict output formatting was required. In text classification, it showed low accuracy for the 'salud' category and struggled with ambiguous or multi-domain texts, scoring 0.69 accuracy on the Sib200 classification SP dataset.

Technical specifications

Context length

8k tokens

Max output

—

Architecture

Transformer

Modalities

text->text

Parameters

~7B

Open weights

Yes

Licence

Gemma

Regions

EU

Vendor

BSC-LT

Provider model key

bsc-lt/salamandra-7b-instruct