CoolFace
Modelpublic

AdaobiU/adaption_llama_4_scout_17b_16_literacy_research_texts_6979ecb8

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face
0likes42downloads
Model Card

Model Card for Igbo Literacy Localisation Model

Model Details

Model Description

This repository contains a PEFT adapter fine-tuned for the localisation of English literacy and education texts into Igbo. The model was developed as part of the AFRICOMPLINGS 2026 Summer School final project and focuses on preserving vowel harmony, tone-marked orthography, and education-domain terminology during localization.

  • —Developed by: Adaobi Ugwuonah
  • —Model type: PEFT adapter for causal language modeling
  • —Language(s): Igbo (ig)
  • —Finetuned from model: togethercomputer/Llama-4-Scout-17B-16E-Instruct_bnb_4bit

Uses

Direct Use

The model can be used to generate or localize educational and literacy-related content in Igbo.

Out-of-Scope Use

The model is not intended for medical, legal, financial, or safety-critical applications. It has not been evaluated for those domains.

Bias, Risks, and Limitations

  • —Tone marking is inconsistent across outputs.
  • —The model frequently preserves English educational terminology through code-switching.
  • —Performance outside the literacy and education domain is unknown.

Training Data

The training corpus consisted of English literacy and education texts collected from publicly available educational resources and localized into Igbo using Adaptionlabs.ai. The corpus included reading instruction, phonics, literacy policy, UNESCO programs, and educational research materials.

Training Procedure

Fine-tuning was performed using the AutoScientist workflow with PEFT adapters.

Evaluation

Metrics

  • —Tone-marking rate
  • —Vowel harmony conformance
  • —Domain-term preservation ratio
  • —chrF style-divergence score

Results

MetricChosenRejected
Tone-marking rate0.83%0.63%
Vowel harmony conformance97.9%97.7%
Avg. domain terms preserved per row2.603.20

Average chrF (chosen vs rejected): 54.37

Summary

The model preserved Igbo vowel harmony with very high fidelity and maintained strong education-domain terminology preservation, while tone-marked orthography remained sparse and inconsistent.

Citation

Ugwuonah, A. (2026). Evaluating Tone Accuracy, Vowel Harmony Fidelity, and Code-Switching Preservation in Automated Localization of English Literacy Texts into Igbo Using Adaptionlabs.ai. AFRICOMPLINGS 2026 Summer School Final Project.

Framework Versions

  • —PEFT 0.15.1