"Affordably Fine-tuned LLMs Provide Better Answers to Course-specific MCQs"

Fine-tuning 7B models with textbooks outsmarts 70B models on course questions.

Fine-tuning smaller LLMs with course textbooks enables them to outperform larger models on Multiple Choice Questions while requiring minimal computing resources.

https://arxiv.org/abs/2501.05891v1

🤔 Original Problem:

→ Educational institutions face challenges using LLMs due to high computational costs and poor performance on domain-specific questions.

-----

🔧 Solution in this Paper:

→ The researchers tested LLaMA-2 variants (7B, 13B, 70B) on 162 Programming Language MCQs.

→ They used LoRA and qLoRA techniques for efficient fine-tuning with course textbook content.

→ The process involved testing different learning rates, batch sizes, and epochs to optimize performance.

-----

💡 Key Insights:

→ Fine-tuned 7B and 13B models can run on consumer GPUs (24GB)

→ Single-chapter fine-tuning produced more stable results than using entire textbook

→ Quantized models showed minimal accuracy loss while reducing memory usage significantly

-----

📊 Results:

→ 13B quantized variants achieved 78% better performance than pre-trained versions

→ Fine-tuned 7B models required only 13GB memory vs 45GB for base models

→ Free tier Google Colab (15GB) can support inference and fine-tuning of 7B quantized models

------

Are you into AI and LLMs❓ Join my daily AI newsletter. I will send you 7 emails a week analyzing the highest signal AI developments. ↓↓

🎉 https://rohanpaul.substack.com/