phi-3-mini-4k-instruct-1it

Model Card: QLoRA Fine‑Tuned Small Language Model

🧩 Model Summary

This model is a QLoRA-fine-tuned variant of an open-source lightweight transformer (Phi-style architecture), trained entirely on Google Colab Free Tier using 4-bit quantization + PEFT to minimize compute and memory.

The project demonstrates how small LLMs can be efficiently adapted, evaluated, and deployed using a fully reproducible low-compute workflow.

---

📁 Repository Structure

/README.md
/config.json
/tokenizer.json
/adapter_config.json
/adapter_model.bin
/training_logs/

---

🛠️ Training Setup

Base Model

• Architecture: Phi-style small transformer
• Precision: 4-bit (NF4)
• Parameters: ~300M (depending on chosen base)

Method

• Fine-tuning: QLoRA (PEFT)
• Optimizer: AdamW
• LR: 2e-4
• Batch Size: 8
• Epochs: 5
• Eval Strategy: Per epoch
• Platform: Google Colab Free Tier (T4 GPU or TPU v2)

---

📊 Evaluation

=== METRIC COMPARISON === ROUGE-L: before=0.2726 after=0.2726 BLEU : before=19.9785 after=19.9744 Perplexity: before=23.67 after=3.02

Graphs (Before → After)

Notebook includes:

• ROUGE-L Bar Chart
• BLEU Bar Chart
• Perplexity Bar Chart

---

📜 Dataset

• Type: Small instruction–response dataset
• Size: ~20–100 examples
• Purpose: Rapid QLoRA testing and demonstration

---

📦 Intended Use

Good for:

• Teaching QLoRA
• Prototyping domain-specific adapters
• Low-compute experimentation
• Research demonstrations

Not intended for:

• High-risk or production tasks
• Safety-critical operations
• Broad general-purpose deployment

---

⚠️ Limitations

• Small dataset (limited generalization)
• No RLHF alignment
• Domain-specific behavior
• Not safety-aligned

---

💻 Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "<your-hf-username>/<your-model-repo>"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")

prompt = "Explain QLoRA in one sentence."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=80)
print(tokenizer.decode(output[0], skip_special_tokens=True))

---

📤 Hugging Face Hub Workflow

The provided notebook automates:

• Authentication
• Repo creation
• Uploading adapters
• Uploading evaluation graphs
• Uploading logs
• Uploading this README.md

---

🧪 Reproducibility

Training workflow:

1. Load base model
2. Apply 4-bit quantization
3. Train QLoRA adapters
4. Evaluate baseline and fine-tuned performance
5. Compare metrics + graphs
6. Push results to HF Hub

Runtime on Google Colab Free Tier: ~40–45 minutes.

---

📚 Credits

• Microsoft — Phi inspiration
• HuggingFace — Transformers, PEFT, Accelerate
• Google Colab — Compute backend
• Project Author — Data, workflow, evaluation

---

📄 License

Follow the base model’s original license (Apache 2.0 / MIT / OpenRAIL / etc.)

---

🤝 Contributions

Pull requests welcome — avoid adding copyrighted or sensitive data.