Mistral Technical Interview QA (LoRA + RAG)

📌 Overview

This model is a fine-tuned version of Mistral-7B-Instruct designed to answer technical interview questions in core Computer Science domains. It is trained on a curated dataset of technical question–answer pairs and evaluated under multiple experimental setups including vanilla LLM, fine-tuning, Retrieval-Augmented Generation (RAG), and hybrid approaches.

The goal of this project is to investigate the effectiveness of parameter-efficient fine-tuning and retrieval augmentation for improving technical question answering in low-resource settings.

🚀 Key Features

Fine-tuned using LoRA (Low-Rank Adaptation) for memory-efficient training
Supports Retrieval-Augmented Generation (RAG)
Focused on Computer Science domains:
- Data Structures and Algorithms
- Operating Systems
- Computer Networks
- Databases
- OOP and Core CS fundamentals
Lightweight adapter weights for fast deployment
Evaluated using BLEU, ROUGE, BERTScore, and Exact Match

🧠 Base Model

Model: mistralai/Mistral-7B-Instruct-v0.2
Architecture: Decoder-only transformer
Training method: Supervised fine-tuning with LoRA

📚 Dataset

The dataset consists of approximately 2,000 technical question–answer pairs covering multiple Computer Science domains.

Dataset preparation:

Initial seed dataset manually curated
Extended using synthetic augmentation
Duplicate and semantic filtering applied
Train / validation / test splits created
Format: JSONL

⚠️ Note: Some data samples were generated using LLM-based augmentation and manually filtered for quality.

⚙️ Training Details

Fine-tuning approach:

Parameter-efficient LoRA tuning
Frozen base model weights
Adapter-only training

Configuration:

LoRA rank: 8
Learning rate: 1e-4
Epochs: 3
Sequence length: 384
Precision: FP16
Gradient accumulation used

Training was performed on an NVIDIA RTX 3090 (24GB VRAM).

🔍 Retrieval-Augmented Generation (RAG)

A semantic retriever was implemented using:

Sentence Transformers (all-MiniLM-L6-v2)
FAISS vector database
Top-k semantic search

Retrieved context is injected into the prompt before generation.

📊 Evaluation

The model was evaluated under four experimental settings:

Vanilla Mistral
RAG + Vanilla
Fine-tuned Mistral
RAG + Fine-tuned

Metrics used:

BLEU-4
ROUGE-L
BERTScore
Exact Match

Results showed that:

Fine-tuning improves answer quality and domain specificity
RAG improves factual grounding
The hybrid approach provides the best performance

🧪 Intended Use

This model is designed for:

Interview preparation
Educational tools
Technical tutoring
Research in retrieval-augmented systems

It is not intended for:

Medical or legal decision-making
Safety-critical systems

⚠️ Limitations

Dataset size is relatively small
Synthetic data may introduce bias
Performance may vary for unseen or highly specialized domains
Retrieval quality depends on the vector database

🔐 Ethical Considerations

The dataset does not contain personally identifiable or sensitive information. The model may generate incorrect or outdated information and should be used with human oversight.

📦 How to Use

Load merged model:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("yourusername/repo-name")
tokenizer = AutoTokenizer.from_pretrained("yourusername/repo-name")

Load LoRA version:

from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
model = PeftModel.from_pretrained(base, "yourusername/repo-name")

📈 Future Work

Larger and more diverse dataset
Multi-hop retrieval
Improved retriever training
Reinforcement learning for answer quality
Human evaluation

hookup

👨‍💻 Author

Final-year Computer Science project focused on applied LLM systems, parameter-efficient fine-tuning, and retrieval-augmented generation.

📜 License

Please refer to the base model license (Mistral AI).