Maloba

Conversational assistant

Llama 3.1 8B + QLoRA

Answer, summarise and explain in Lingala, across five registers.

Merged weightsAdaptersTraining report F01

Result

ROUGE-L 0,2379 contre 0,1521 pour la base on 964 test examples. Source : E01. Voir la page Résultats.

Mesures automatiques (E01)
MesureValeur
ROUGE-L0,2379
ROUGE-L, modèle de base0,1521

The measurements, as charts

Base model against fine-tuned modelROUGE-L measures overlap with the expected answer. The gain is clear, but no external reference exists for Lingala on this task.
ROUGE-L0,15210,2379+0,0858 · p —Modèle de baseModèle affiné Maloba

Source : E01, 964 exemples de test

Show the chart data
MeasureBaseFine-tunedGap
ROUGE-L0,15210,2379+0,0858

What it does well, what it does badly

What it does well

  • ROUGE-L 0.2379 against 0.1521 for the base model: +56 %.
  • Five registers learned: translation, formal, urban, teaching, summary.
  • Merged model and adapters published.

What it does badly

  • An assistant in a low-resource language stays fragile: check its answers before use.
  • The training corpus has a strict human validity of 58 % (weak agreement).
  • Llama 3.1 Community License: use is subject to its terms.

How it was built

Socle meta-llama/Llama-3.1-8B-Instruct, adapté par qLoRA, LoRA rang 16. Données : 5 700 exemples Alpaca, 5 registres. Détail dans le rapport F01.

Licence

Llama 3.1 Community License : subject to licence terms. Un modèle garde la licence de son modèle de base, quelle que soit celle des données d'entraînement. Gouvernance des données.