| --- |
| base_model: unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit |
| library_name: peft |
| pipeline_tag: text-generation |
| tags: |
| - qwen2.5 |
| - lora |
| - qlora |
| - unsloth |
| - peft |
| - transformers |
| - trl |
| - personal-ai |
| - digital-twin |
| - personality-model |
| - text-generation |
| --- |
| |
| # MeetMe V1 - Personal AI |
|
|
| MeetMe V1 is a LoRA fine-tuned version of Qwen2.5-3B-Instruct designed to replicate my personal writing style, reasoning process, communication style, and decision-making patterns. The model is intended to act as a personal AI companion rather than a general-purpose chatbot. |
|
|
| --- |
|
|
| # Model Details |
|
|
| ## Model Description |
|
|
| MeetMe V1 is a supervised fine-tuned (SFT) language model built using QLoRA on top of Qwen2.5-3B-Instruct. |
|
|
| Instead of teaching new knowledge, the objective of this fine-tuning is personality transfer. The model learns how I communicate, explain concepts, reason through problems, describe experiences, and express opinions while retaining the broad knowledge of the original Qwen model. |
|
|
| The first version was trained using approximately 171 curated conversational samples converted into ChatML format. |
|
|
| --- |
|
|
| ### Developed by |
|
|
| Sanskar |
|
|
| ### Model Type |
|
|
| Causal Language Model (Decoder-only Transformer) |
|
|
| ### Base Model |
|
|
| unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit |
|
|
| ### Fine-tuning Method |
|
|
| QLoRA |
|
|
| ### Framework |
|
|
| Unsloth |
|
|
| ### Language |
|
|
| English |
|
|
| ### Intended Domain |
|
|
| Personal Assistant |
|
|
| Digital Twin |
|
|
| Reasoning Assistant |
|
|
| Learning Companion |
|
|
| ### License |
|
|
| Apache 2.0 (inherits from the base model) |
|
|
| --- |
|
|
| # Model Sources |
|
|
| Base Model |
|
|
| https://huggingface.co/unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit |
|
|
| Framework |
|
|
| https://github.com/unslothai/unsloth |
|
|
| Qwen |
|
|
| https://github.com/QwenLM |
|
|
| --- |
|
|
| # Intended Uses |
|
|
| ## Direct Use |
|
|
| MeetMe V1 is designed to: |
|
|
| - answer questions using my communication style |
| - explain technical concepts the way I usually explain them |
| - simulate my reasoning process |
| - act as my personal AI assistant |
| - assist in brainstorming and project planning |
| - answer questions about my learning philosophy and opinions |
|
|
| --- |
|
|
| ## Downstream Use |
|
|
| Possible applications include: |
|
|
| - Personal chatbot |
| - AI mentor |
| - Digital twin |
| - Personal journaling assistant |
| - Memory retrieval system |
| - AI-powered second brain |
| - Personal productivity assistant |
|
|
| --- |
|
|
| # Out-of-Scope Uses |
|
|
| This model should not be used as: |
|
|
| - Medical advisor |
| - Legal advisor |
| - Financial advisor |
| - Psychological counselor |
| - Source of factual truth |
| - Identity verification system |
|
|
| The model reflects my personal experiences and opinions and should not be treated as objective knowledge. |
|
|
| --- |
|
|
| # Training Data |
|
|
| The dataset consists of manually curated personal conversations and journal entries. |
|
|
| Dataset characteristics: |
|
|
| - Approximately 171 conversation samples |
| - ChatML format |
| - User → Assistant conversations |
| - Human-written responses |
| - Categories include: |
| - AI |
| - Programming |
| - Learning |
| - Career |
| - Productivity |
| - Relationships |
| - Personal goals |
| - Habits |
| - Decision making |
| - Philosophy |
|
|
| The dataset focuses on preserving personality rather than teaching factual knowledge. |
|
|
| --- |
|
|
| # Data Preprocessing |
|
|
| Before training, the data underwent: |
|
|
| - Manual cleaning |
| - Grammar correction where necessary |
| - Duplicate removal |
| - ChatML conversion |
| - Validation |
| - JSON formatting |
| - Quality filtering |
|
|
| --- |
|
|
| # Training Procedure |
|
|
| The model was fine-tuned using QLoRA through Unsloth. |
|
|
| ## Training Configuration |
|
|
| Base Model |
|
|
| Qwen2.5-3B-Instruct |
|
|
| Method |
|
|
| QLoRA |
|
|
| Precision |
|
|
| 4-bit Quantization |
|
|
| Optimizer |
|
|
| AdamW 8-bit |
|
|
| Framework |
|
|
| Unsloth |
|
|
| PEFT |
|
|
| LoRA |
|
|
| --- |
|
|
| # LoRA Configuration |
|
|
| Rank (r) |
|
|
| 16 |
|
|
| Alpha |
|
|
| 16 |
|
|
| Dropout |
|
|
| 0 |
|
|
| Target Modules |
|
|
| - q_proj |
| - k_proj |
| - v_proj |
| - o_proj |
| - gate_proj |
| - up_proj |
| - down_proj |
| |
| --- |
| |
| # Training Hyperparameters |
| |
| Epochs |
| |
| 3 |
| |
| Learning Rate |
| |
| 2e-4 |
| |
| Sequence Length |
| |
| 2048 |
| |
| Packing |
| |
| Enabled |
| |
| Batch Size |
| |
| 2 |
| |
| Gradient Accumulation |
| |
| 4 |
| |
| Scheduler |
| |
| Linear |
| |
| Weight Decay |
| |
| 0.01 |
| |
| Warmup Ratio |
| |
| 0.03 |
| |
| Random Seed |
| |
| 3407 |
| |
| --- |
| |
| # Evaluation |
| |
| Evaluation focused on personality consistency instead of benchmark scores. |
| |
| Evaluation criteria included: |
| |
| - Personality preservation |
| - Writing style similarity |
| - Decision-making consistency |
| - Reasoning quality |
| - Emotional consistency |
| - Generalization to unseen questions |
| |
| The model demonstrated good personality transfer despite being trained on a relatively small dataset. |
| |
| --- |
| |
| # Known Limitations |
| |
| - Limited training dataset (171 samples) |
| - May answer outside my personality on unseen domains |
| - Can still inherit behaviors from the base Qwen model |
| - Does not possess persistent memory |
| - Does not know events occurring after the dataset was created |
| |
| --- |
| |
| # Future Improvements |
| |
| Planned improvements include: |
| |
| - Increase dataset to over 1000 curated conversations |
| - Add Retrieval-Augmented Generation (RAG) |
| - Long-term memory |
| - Multi-turn conversation training |
| - Continuous incremental fine-tuning |
| - Emotion-aware responses |
| - Voice cloning integration |
| |
| --- |
| |
| # Technical Specifications |
| |
| Architecture |
| |
| Transformer Decoder |
| |
| Base Parameters |
| |
| 3 Billion |
| |
| Fine-tuning |
| |
| QLoRA Adapter |
| |
| Inference |
| |
| Transformers + PEFT |
| |
| Framework |
| |
| PyTorch |
| |
| --- |
| |
| # Hardware |
| |
| Google Colab |
| |
| NVIDIA Tesla T4 GPU |
| |
| --- |
| |
| # Software |
| |
| Python |
| |
| PyTorch |
| |
| Transformers |
| |
| Datasets |
| |
| PEFT |
| |
| TRL |
| |
| Unsloth |
| |
| BitsAndBytes |
| |
| --- |
| |
| # Citation |
| |
| If you use MeetMe V1 in your research or projects, please cite this repository. |
| |
| ```bibtex |
| @software{meetme_v1, |
| author = {Sanskar}, |
| title = {MeetMe V1: A Personal AI Digital Twin}, |
| year = {2026}, |
| url = {https://github.com/<your-github-username>/MeetMe} |
| } |
| ``` |
| |
| --- |
| |
| # Project Vision |
| |
| MeetMe aims to create a digital twin capable of preserving personality, reasoning style, communication patterns, and long-term personal knowledge. Rather than replacing human decision-making, the goal is to augment memory, productivity, and learning through a personalized AI assistant that evolves alongside its creator. |
