qwen-personality-lora

Qwen 1.5B LoRA Fine-Tuned for Personality Conditioning

Model Description

This model is a parameter-efficient fine-tuned version of Qwen2.5-1.5B-Instruct, adapted to demonstrate controllable personality-driven behavior in instructional responses.

The model was fine-tuned using synthetic instruction–response pairs generated in a KnowSLM-style format. Each instruction explicitly specifies a personality trait that conditions the tone and structure of the response.

The primary objective is to enable behavioral control while preserving technical correctness and instructional clarity.

---

Base Model

• Base Model: Qwen2.5-1.5B-Instruct
• Quantization: 4-bit NF4
• Fine-tuning Method: LoRA (Low-Rank Adaptation)

---

Training Objective

To condition the model’s response style based on specified personality traits while maintaining technical correctness and instructional clarity.

---

Personality Traits

The dataset consists of 100 synthetic instruction–response samples distributed across four personality traits:

• Openness (High)
• Openness (Low)
• Conscientiousness (High)
• Agreeableness (High)

Each instruction explicitly specifies the personality trait to guide response tone and structure.

---

Training Method

LoRA was applied to the following attention projection layers:

• q_proj
• k_proj
• v_proj

Training Configuration

• Trainable Parameters: 0.0965% of total parameters
• Learning Rate: 1e-4
• Epochs: 3
• Batch Size: 2
• Gradient Accumulation: 4
• Max Sequence Length: 512
• Training Device: Google Colab GPU
• Precision: 4-bit quantized (NF4)

---

Dataset

• Size: 100 synthetic instruction–response pairs
• Format: JSONL

Structure

Each sample contains:

• instruction
• input
• output

Topic Domains

• Retrieval-Augmented Generation (RAG)
• Embeddings
• LoRA
• Quantization
• Vector Search
• Access Control

---

Evaluation

Evaluation was conducted qualitatively to assess tone variation across personality prompts.

Observed Behavior

• Openness (High): Creative and exploratory responses
• Openness (Low): Cautious and traditional tone
• Conscientiousness (High): Structured, step-by-step answers
• Agreeableness (High): Warm and supportive tone

Training Loss

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

Loss reduction indicates effective adaptation.

---

Intended Use

This model demonstrates:

• Personality-conditioned instruction tuning
• Parameter-efficient fine-tuning
• Behavioral control in large language models

It is intended for research and educational purposes.

---

Limitations

• Small dataset (100 samples)
• Synthetic training data
• Primarily qualitative evaluation
• Not production-ready

---

Citation

If referencing this model in academic work, please cite:

Qwen2.5-1.5B-Instruct (Alibaba Qwen Team)

---

Author

Samiksha Jangid
Agentic AI – Assignment 2