| --- |
| license: cdla-sharing-1.0 |
| base_model: microsoft/phi-2 |
| tags: |
| - customer-support |
| - phi-2 |
| --- |
| |
| # Xenia — Customer Support AI |
|
|
| Named after Xenia, the ancient Greek concept of hospitality and guest-friendship. |
|
|
| Fine-tuned from microsoft/phi-2 to generate natural, helpful customer support responses. |
|
|
| ## Intended Use |
| Generates support responses for order tracking, cancellations, refunds, payment issues, and delivery changes. |
|
|
| ## Training Data |
| Fine-tuned on bitext/Bitext-customer-support-llm-chatbot-training-dataset (26,872 rows, license: cdla-sharing-1.0). |
|
|
| ## Prompt Format |
| Instruct: {your question} |
| Output: |
|
|
| ## Placeholders |
| Responses may contain placeholders like {{Order Number}} — intentional, replace with real values before showing to end users. |
|
|
| ## Loading |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| model = AutoModelForCausalLM.from_pretrained('sibte-dev/xenia') |
| tokenizer = AutoTokenizer.from_pretrained('sibte-dev/xenia') |
|
|
| ## Limitations |
| - May hallucinate details not present in the input. |
| - English-only. |
| - Not evaluated on adversarial/abusive input. |
|
|
| ## Training Details |
| - Base: microsoft/phi-2 (2.7B, MIT) |
| - Method: LoRA (r=8, alpha=16, target: q_proj, k_proj, v_proj), then merged |
| - Trainable params during fine-tune: 3.9M (0.14%) |
