qwen-personality-lora

Qwen2.5-1.5B Personality-Conditioned LoRA Model

Overview

This repository contains a parameter-efficient adaptation of Qwen2.5-1.5B-Instruct, fine-tuned to demonstrate controllable personality-aware instructional behavior.

The model has been adapted using LoRA (Low-Rank Adaptation) to condition response tone based on explicitly specified personality traits while preserving factual consistency and technical clarity.

The objective of this project is to explore lightweight behavioral steering in large language models using synthetic supervision.

---

Base Architecture

• Foundation Model: Qwen2.5-1.5B-Instruct
• Model Size: ~1.5 Billion parameters
• Quantization: 4-bit NF4 (BitsAndBytes)
• Fine-Tuning Strategy: LoRA (Low-Rank Adaptation)

Only a small fraction of parameters (~0.0965%) were updated, making the training process computationally efficient and suitable for limited hardware environments.

---

Project Objective

The primary goals of this fine-tuning process were to:

• Introduce personality-driven stylistic variation
• Maintain instructional and technical correctness
• Demonstrate efficient behavioral conditioning using minimal data
• Validate LoRA as a lightweight personality adaptation mechanism

Rather than modifying core model knowledge, the training focused on adjusting stylistic and structural response tendencies.

---

Personality Conditioning Framework

The dataset includes synthetic instruction-response pairs categorized across four behavioral profiles:

• High Openness
• Low Openness
• High Conscientiousness
• High Agreeableness

Each prompt explicitly includes a personality directive such as:

"You are a teacher with personality trait: [Trait Name]"

This explicit conditioning enables controlled variation in tone, structure, and delivery style.

---

Fine-Tuning Methodology

LoRA adapters were applied to selected attention projection modules:

• q_proj
• k_proj
• v_proj

These layers were targeted because attention projections strongly influence contextual emphasis and response formulation, making them suitable for stylistic adaptation.

---

Training Configuration

• Trainable Parameters: ~0.0965%
• Learning Rate: 1e-4
• Epochs: 3
• Batch Size: 2
• Gradient Accumulation Steps: 4
• Maximum Sequence Length: 512
• Hardware Used: Google Colab GPU
• Precision Mode: 4-bit NF4 quantization

The setup ensured efficient resource utilization while maintaining stable convergence.

---

Dataset Details

• Total Samples: 100 synthetic instruction-response pairs
• Format: JSONL

Structure Fields

• instruction
• input
• output

Covered Technical Domains

• Retrieval-Augmented Generation (RAG)
• Vector Embeddings
• LoRA Fine-Tuning
• Quantization Techniques
• Vector Search
• Access Control Systems

The synthetic design ensures controlled exposure to personality-driven tone modulation across consistent technical contexts.

---

Training Performance

Loss Trend

• Initial Loss: ~5.8
• Final Loss: ~2.05

The steady reduction in loss indicates successful adaptation of the LoRA layers without destabilizing the base model.

---

Behavioral Evaluation

Evaluation was conducted qualitatively by prompting identical technical queries under different personality instructions.

Observed Response Patterns

• High Openness: Expansive, idea-driven explanations encouraging exploration.
• Low Openness: Conservative, traditional framing with minimal deviation.
• High Conscientiousness: Structured, logical, step-wise responses.
• High Agreeableness: Supportive, empathetic, learner-friendly tone.

Compared to the base model’s neutral output, the fine-tuned version shows measurable stylistic divergence aligned with the requested trait.

---

Intended Applications

This model is designed for:

• Research on controllable LLM behavior
• Demonstrations of parameter-efficient fine-tuning
• Educational experimentation with personality-driven AI systems
• Study of stylistic modulation in instruction-tuned models

This is a research prototype and not intended for production deployment.

---

Limitations

• Limited dataset size (100 samples)
• Fully synthetic supervision
• Evaluation primarily qualitative
• No benchmarking on factual accuracy improvements
• Personality conditioning may not generalize outside training domains

---

Attribution

If referencing this work, please acknowledge:

• Base Model: Qwen2.5-1.5B-Instruct (Alibaba Qwen Team)

---

Author

Kartik Mehra
Agentic AI – Assignment 2