kevembuvak/llama32-1b-qlora-finetome-exam-gguf

ID2223 Exam Coach – QLoRA-Fine-Tuned Llama 3.2 1B (GGUF)

This repository contains a 1B parameter Llama 3.2 Instruct model fine-tuned using QLoRA in two stages:

1. Instruction tuning on the FineTome dataset
2. Exam-specialized fine-tuning on a custom ID2223 exam preparation dataset

• The final model is exported to GGUF format (Q4_K_M and Q8_0) for efficient CPU inference.
• It powers the Hugging Face Space "kevembuvak/iris", a self-study assistant for the KTH course “ID2223 – Scalable Machine Learning and Deep Learning”.

---

Model Summary

Property	Value
Base Model	meta-llama/Llama-3.2-1B-Instruct
Fine-tuning method	QLoRA (4-bit NF4 quantization)
Training Frameworks	transformers, trl, peft, bitsandbytes
GPU Used	NVIDIA RTX 4090 16 GB
LoRA Config	r=16, alpha=32, dropout=0.05
Context length	4096
Export Format	GGUF Q4_K_M and Q8_0
Intended Use	ID2223 exam preparation, ML/LLM explanations

---

Training Pipeline

This model was trained in two QLoRA stages.

Stage 1 - Instruction Tuning (FineTome)

Dataset: mlabonne/FineTome-100k (sampled 40k)

Preprocessing steps:

• Convert ShareGPT-style data to Llama-3 message format
• Apply chat template (tokenizer.apply_chat_template)
• Limit to 4 most recent dialogue turns
• Truncate to max length 512 tokens
• Create 99/1 train-validation split

Training configuration:

• 2 epochs
• LR = 2e-4
• Batch size 4 × grad_accum 4
• BF16 compute
• Optimizer: paged_adamw_8bit

This produced the adapter: llama32-1b-qlora-finetome.

---

Stage 2 - ID2223 Exam Specialization

Dataset: kevembuvak/id2223_exam_prep

Training configuration:

• Base model: Meta-Llama-3.2-1B-Instruct
• Load Stage 1 adapter as trainable
• LR = 1e-4
• 2 epochs
• Batch size 4 × grad_accum 4
• BF16 compute
• Optimizer: paged_adamw_8bit

Final training loss: approx. 0.079

---

Related Resources

Companion Hugging Face Space - Iris [https://huggingface.co/spaces/kevembuvak/iris]

Training Code (GitHub) [https://github.com/keremburakyilmaz/PEFT-of-LLM]

---

Intended Use & Limitations

Intended For:

• ID2223 exam preparation
• Generating study materials and quizzes
• Evaluating written answers

Not Suitable For:

• High-stakes or safety-critical tasks
• Factual-heavy real-world use outside ML context

Acknowledgements

• Meta AI for Llama 3.2

• FineTome dataset (mlabonne/FineTome-100k)

• Hugging Face (transformers, trl, peft)