phi3-mini-support-bot

Phi-3-mini-4k-instruct — Customer Support Fine-tune

A LoRA fine-tune of Microsoft's Phi-3-mini-4k-instruct (3.8B parameters), adapted for retail/e-commerce customer support conversations. Trained using Unsloth with 4-bit QLoRA for fast, memory-efficient fine-tuning on a single T4 GPU.

Model Details

• Base model: unsloth/Phi-3-mini-4k-instruct
• Fine-tuning method: QLoRA (4-bit), rank 16, alpha 16
• Framework: Unsloth + TRL SFTTrainer
• Training hardware: 1x NVIDIA T4 (Google Colab, free tier)
• Training data: 467 cleaned, deduplicated customer support Q&A pairs (custom dataset)

Training Data

The training set covers common retail support scenarios across these categories:

Each example is a single user question paired with a support-style answer (e.g. order tracking, return policy, payment troubleshooting). Data was cleaned to fix text-encoding artifacts and deduplicated before training.

Intended Use

This model is intended for retail customer-support-style Q&A within the categories above — e.g. answering questions about order status, return policy, shipping timelines, and payment issues in a tone consistent with the training data.

It is not intended for:

• General-purpose assistant use outside customer support
• Safety-critical or compliance-sensitive support (e.g. legal, medical, financial advice)
• Production deployment without human review of outputs

Limitations

• Trained on a small dataset (467 examples) — behavior on questions outside the categories above is closer to the unmodified base model.
• Some categories (Sizing, Order Management, Support) have very few examples and may show limited improvement over the base model.
• Like any fine-tuned language model, it can produce plausible-sounding but incorrect specifics (order numbers, policy details, dates) — outputs should be verified before being shown to real customers.
• English only.

How to Use

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained("00harshh/phi3-mini-support-bot")
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content": "Where is my order?"}]
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
).to("cuda")

outputs = model.generate(input_ids=inputs, max_new_tokens=150, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Training Procedure

• LoRA config: r=16, alpha=16, dropout=0, target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Optimizer: adamw_8bit
• Learning rate: 2e-4, linear schedule
• Epochs: 3
• Effective batch size: 16 (batch size 4 × gradient accumulation 4)