mzansilm-tutor

MzansiLM-Tutor

MzansiLM-Tutor is a specialized instruction-tuned version of anrilombard/mzansilm-125m, specifically designed to act as an AI Tutor for the South African CAPS (Curriculum and Assessment Policy Statement) curriculum.

Model Details

• Model Name: MzansiLM-Tutor
• Base Model: anrilombard/mzansilm-125m
• Architecture: Llama-based architecture (125M parameters)
• Languages: English, Afrikaans, isiZulu, isiXhosa, Sesotho, Sepedi, Setswana, Xitsonga, Siswati, isiNdebele, Tshivenda
• License: MIT

Intended Uses & Limitations

Intended Uses

This model is intended to be used as an educational assistant and tutor. It excels at:

• Explaining core concepts from the South African curriculum (Grade 1-12).
• Generating multiple-choice questions (MCQs), flashcards, and interactive lesson templates.
• Assisting students in their mother tongue across all 11 official South African languages.
• Summarizing textbook materials into structured study formats.

Limitations & Ethical Considerations

• Size Constraints: As a 125M parameter model, it may hallucinate or struggle with highly complex reasoning tasks compared to larger frontier models (like GPT-4 or Claude).
• Curriculum Scope: Its knowledge is heavily anchored in the CAPS curriculum. It may not perform optimally outside of these specific educational boundaries.
• Language Nuance: While trained on all 11 official languages, proficiency may vary depending on the specific language's representation in the training dataset.

Training Details

Dataset

The model was fine-tuned on a proprietary dataset containing 268,794 instruction-tuning examples. This dataset was built by aggregating:

• DBE (Department of Basic Education) National Catalogues and Workbooks.
• Translated curriculum topics, descriptions, and learning objectives across 11 languages.
• Thousands of AI-generated question templates, flashcards, and lesson summaries.

Training Procedure

• Hardware: Fine-tuned on RunPod GPUs.
• Method: Parameter-Efficient Fine-Tuning (PEFT) using LoRA (Low-Rank Adaptation) and then merged back into the base weights.
• Steps: 500 steps
• Final Training Loss: 2.101
• Precision: float16

Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "misterseitz/mzansilm-tutor"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    low_cpu_mem_usage=True
)

prompt = "SYSTEM: You are a Grade 10 Physical Sciences tutor.\nSTUDENT: Explain Newton's First Law of Motion.\nTUTOR:"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Acknowledgments

• Base Model: We acknowledge Anri Lombard for the creation of the foundational MzansiLM model.
• Curriculum Data: The Department of Basic Education (DBE) for their open-access National Catalogues.