SinLlama-MCQ is a specialized fine-tuned version of Llama-3-8B designed to generate Multiple Choice Questions (MCQs) in Sinhala. This model was developed to assist educators and students in Sri Lanka by automating the creation of high-quality assessment materials.

Model Details

Developed by: Joseph Rodrigo
Model Type: PeftAdapter (LoRA)
Base Model: meta-llama/Meta-Llama-3-8B
Language(s): Sinhala (Primary), English
Task: Causal Language Modeling / MCQ Generation

Technical Specifications

LoRA Hyperparameters

The model was trained using the following configuration:

Rank (r): 64
Alpha (lora_alpha): 128.0
Dropout (lora_dropout): 0.05
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Modules to Save: embed_tokens, lm_head (to ensure Sinhala script compatibility)

Framework Versions

PEFT: 0.18.1
Transformers: Latest compatible with Llama-3

Intended Use

Direct Use

This model is intended for:

Generating 4-option MCQs from Sinhala text paragraphs.
Creating educational content for Sri Lankan schools and private tutors.

Out-of-Scope Use

General-purpose chat not related to educational assessments.
Critical medical or legal advice without human review.

How to Load (FastAPI / Python)

To use this model in your backend, ensure you have the peft and transformers libraries installed.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_path = "meta-llama/Meta-Llama-3-8B"
adapter_path = "YOUR_HF_USERNAME/sinllama-mcq"

tokenizer = AutoTokenizer.from_pretrained(base_model_path)
model = AutoModelForCausalLM.from_pretrained(
    base_model_path, 
    torch_dtype=torch.float16, 
    device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_path)

def generate_mcq(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    with torch.no_grad():
        outputs = model.generate(**inputs, max_new_tokens=512)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

itsjorigo/sinllama-mcq

SinLlama-MCQ