📚 Academic Style Tuning for LLMs

This project explores the fine-tuning of large language models (LLMs), specifically GPT-2 and Mistral-7B, to better emulate formal academic writing. Using a curated subset of arXiv abstracts, we demonstrate how domain-specific fine-tuning can enhance stylistic fidelity in scholarly communication—though limitations like hallucinations and generic phrasing remain.

👨‍💻 Team Members

• Joshua Sun (zysun@ucdavis.edu)
• Keer Ni (knni@ucdavis.edu)
• Muhammad Reza (msreza@ucdavis.edu)

📁 Repository Structure

• 🗂️ Main Project: [This repo]
• 🔬 GPT-2 Fine-Tuning: Hugging Face
• 🔬 Mistral-7B Fine-Tuning (QLoRA): Hugging Face

🔍 Motivation

While LLMs such as GPT-2 are powerful, they often produce text that lacks the tone, structure, and clarity expected in academia. Our goal was to improve stylistic alignment for use in academic writing assistants, research tools, and educational applications.

🧠 Models

Model	Method	Params	Fine-Tuning Technique
GPT-2 (124M)	Full training	124M	Cross-entropy loss
Mistral-7B	QLoRA	7B	Causal LM with prompts

Both models were fine-tuned on ~500K abstracts from the arXiv dataset.

🛠️ Methodology

• Preprocessing: Abstract cleaning, LaTeX removal, length standardization, GPT-2 BPE tokenization.
• Training:
  • GPT-2: Plain-text format, trained for 1.2 epochs.
  • Mistral-7B: Prompt-target pairs (title → abstract), trained with QLoRA for 1 epoch.
• Evaluation: Human-centered evaluation focused on academic tone, fluency, and coherence.

📈 Results

• GPT-2 showed improvement in tone but struggled with coherence and syntax.
• Mistral-7B was more fluent but often produced generic or shallow content.
• Both models suffered from hallucinations and logical inconsistencies.

📊 Results Summary

Metric	GPT-2 (Full Tune)	Mistral-7B (QLoRA)
Academic Tone	✅ Moderate	✅ Strong
Coherence	⚠️ Inconsistent	✅ Mostly Fluent
Hallucinations	❌ Frequent	❌ Still present
Specificity / Depth	❌ Shallow	⚠️ Generic

💡 Key Insights

• Light fine-tuning alone is insufficient for high-quality academic generation.
• Parameter-efficient methods like QLoRA can adapt larger models with limited resources.
• Evaluation must be human-centered to assess stylistic and logical performance.

🧪 Quick Example

Prompt:
"LoRA"

GPT-2 Output (sample):
"LoRA) method to solve the linear problem of constructing a convex hull for an arbitrary non-degenerate group , in this paper we propose and prove that, under certain conditions on , such as the dimension of its manifold with respect (or ), the -norm is bounded above by ."

Mistral Output (sample):
"We present LoRA (Loosely Regularized Adapters), a novel approach to fine-tune large language models (LLMs) for in-context learning tasks. Unlike conventional fine-tuning methods that modify the model's weights, LoRA only adds a few hundred additional trainable parameters to the model, reducing the risk of overfitting and maintaining the model's core functionality."

🚀 Future Work

• Use full 2.7M arXiv dataset
• Experiment with more powerful base models (e.g., Mixtral, Deepseek)
• Introduce retrieval-augmented generation and citation grounding

📎 Report

👉 Final Project Report (PDF)

📽️ Presentation Slides

👉 Final Presentation (PDF)