PAVLY112/xevara-voice-lora

Xevara Voice LoRA

A LoRA (Low-Rank Adaptation) fine-tune of Qwen2.5-7B-Instruct trained to speak consistently in Xevara's voice — the tone, framing, and personality of an AI agency's assistant — without needing a long instruction prompt every time.

This is a genuine fine-tuning project, not a prompted persona: the model's own weights were adjusted through supervised fine-tuning on a custom dataset, using LoRA for efficient training on a free Google Colab T4 GPU via the Unsloth library.

🔗 See it in action: live side-by-side comparison demo — ask the base model and this fine-tuned model the same question and see the difference in tone directly.

Model Details

Model Description

• Developed by: Pavly Esmat (PAVLY112)
• Model type: Causal language model, LoRA adapter
• Language(s): English
• License: Apache 2.0 (inherited from the base model)
• Finetuned from model: unsloth/Qwen2.5-7B-Instruct-bnb-4bit

Model Sources

• Portfolio project page: Xevara — Fine-Tuned Model Comparison
• Training notebook: Google Colab (LoRA fine-tuning via Unsloth)

Uses

Direct Use

This adapter is intended as a portfolio demonstration of custom model fine-tuning. It answers questions about a fictional AI agency ("Xevara") in a warm, confident, concise tone — no filler phrases like "As an AI...", no scripted or generic-sounding responses.

It is not intended for production deployment as-is; it's a proof of concept showing that a model's underlying behavior can be shaped through training rather than prompting alone.

Out-of-Scope Use

Not intended for factual question-answering outside the Xevara persona/domain, safety-critical applications, or any use case requiring guaranteed factual accuracy — like any small fine-tuned model, it can still produce inaccurate or inconsistent responses outside its trained scope.

Bias, Risks, and Limitations

This model was fine-tuned on a small, hand-curated dataset (~80 examples) focused on a narrow persona and topic area. It has not been evaluated for bias, safety, or robustness beyond that scope, and inherits any limitations present in the base Qwen2.5-7B-Instruct model.

How to Get Started with the Model

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="PAVLY112/xevara-voice-lora",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "You are Xevara's assistant."},
    {"role": "user", "content": "What does Xevara do?"},
]
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, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Training Details

Training Data

A custom dataset of ~80 hand-written question/answer pairs covering the fictional Xevara AI agency's services, FAQ-style questions, differentiation from generic chatbots, and small talk — all written in a consistent warm, confident, concise tone.

Training Procedure

Fine-tuned using LoRA via Unsloth on a free Google Colab T4 GPU.

Training Hyperparameters

• LoRA rank (r): 16
• LoRA alpha: 16
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Epochs: 3
• Learning rate: 2e-4
• Optimizer: adamw_8bit
• Quantization: 4-bit (QLoRA-style training)
• Training regime: Mixed precision (fp16/bf16, auto-detected)

Framework Versions

• PEFT
• TRL (SFTTrainer)
• Unsloth

---

Part of the Xevara portfolio project — an AI agency concept demonstrating custom AI system development, from chatbots and autonomous agents to multi-agent pipelines and genuine model fine-tuning.