Congo-digital-service/Llama-3.1-8B-Instruct-Lingala-QLoRA-merged-v2

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Jul 27, 2026License:llama3.1Architecture:Transformer Gated Featherless Exclusive Cold

Congo-digital-service/Llama-3.1-8B-Instruct-Lingala-QLoRA-merged-v2 is an 8 billion parameter instruction-tuned causal language model, built upon Meta's Llama 3.1-8B-Instruct architecture. Developed by Congo Digital Services (CDS SARL), this model is specifically fine-tuned using QLoRA for conversational text generation in Lingala across five distinct registers. It excels at producing fluent Lingala text for urban, educational, summarization, formal, and translation contexts, achieving a ROUGE-L F1 score of 0.2379 on its instruction test set.

Loading preview...

Model Overview

Congo-digital-service/Llama-3.1-8B-Instruct-Lingala-QLoRA-merged-v2 is an 8 billion parameter language model developed by Congo Digital Services (CDS SARL) for the MALOBA project. It is a supervised fine-tune (using QLoRA) of Meta's meta-llama/Llama-3.1-8B-Instruct base model, specifically adapted for the Lingala language. The QLoRA adapters have been merged into the base weights, resulting in a standalone 8B parameter model.

Key Capabilities

  • Lingala Conversational Generation: Optimized for generating conversational text in Lingala.
  • Five Registers: Capable of producing text in urban, educational, summarization, formal, and translation registers.
  • Performance: Achieved a ROUGE-L F1 score of 0.2379 on its 964-example held-out instruction test set, representing a +56.4% relative improvement over the base model.
  • Deployment: This model is served by the MALOBA project under the alias maloba-llm.

Limitations and Considerations

  • Instruction Following: The model primarily learned to write in five registers rather than following varied instructions, due to the nature of its training data.
  • Translation Performance: While it has a translation register, its French → Lingala translation performance (BLEU 1.27) is significantly lower than dedicated translation models like NLLB-200 1.3B (BLEU 22.96).
  • Factual Accuracy: Not recommended for tasks requiring high factual accuracy (e.g., medical, legal) without human review.
  • Orthography: Trained before corpus harmonization, it may reproduce inconsistent spellings.

Use Cases

  • Direct Use: Generating conversational Lingala text using transformers or compatible inference servers.
  • MALOBA Project: The primary LLM for the MALOBA project's Lingala text generation needs.

Finetunes
1 models