papasega/qwen3-wolof-lora

Wolof LoRA Adapter

This repository contains a LoRA adapter fine-tuned for a classroom demo on Wolof instruction following in Senegalese/African contexts.

Base Model

• Qwen/Qwen3-0.6B

Intended Use

This adapter is designed for teaching:

• instruction fine-tuning,
• LoRA deployment,
• local inference,
• basic evaluation with Exact Match, F1, BLEU, ROUGE-L, and perplexity.

It is not a production Wolof assistant.

Loading Example

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "Qwen/Qwen3-0.6B"
adapter_id = "YOUR_USERNAME/YOUR_REPO"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(base_model, trust_remote_code=True)
model = PeftModel.from_pretrained(base, adapter_id)

Training Setup

• LoRA rank: 16
• LoRA alpha: 32
• Target modules: attention and MLP projection layers
• Dataset schema: instruction, input, output
• Chat template rendered without hidden thinking traces when supported.

Limitations

The dataset is small and classroom-oriented. The model may repeat short Wolof phrases or fail outside the covered categories. Evaluate before reuse.