PabloBaeza/tinyllama-professional-etiquette-adapter

TinyLlama Professional Etiquette Assistant (LoRA Adapter)

Model Description

This model is a QLoRA fine-tuned adapter for TinyLlama-1.1B-Chat-v1.0. It has been specifically trained to act as a Professional Etiquette Expert. Its primary task is to transform rude, aggressive, or informal sentences into professional, polite, and corporate-appropriate communication.

• Developed by: PabloBaeza
• Model type: Causal Language Model (LoRA Adapter)
• Base Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
• Task: Sentence Rephrasing / Professional Etiquette

Intended Use

The model is designed for corporate environments (e.g., Slack, Email, Customer Support) where technical staff or users need to rephrase direct or blunt messages into formal language.

Example:

• Input: "Shut up and do your job."
• Output: "I need to ensure that we’re all working towards our goals."

Training Procedure

The model was trained using QLoRA (4-bit quantization) on a Tesla T4 GPU (Google Colab).

Hyperparameters:

• Rank (r): 16
• Alpha: 32
• Target Modules: q_proj, v_proj, k_proj, o_proj
• Learning Rate: 2e-4
• Steps: 100
• Batch Size: 4

Training Results:

• Initial Loss: 1.79
• Final Loss: 0.45
• Convergence: The model showed steady convergence over 100 steps, indicating successful style adaptation.

Limitations

Due to its small size (1.1B parameters), the model may occasionally repeat parts of the input if the repetition penalty is not set correctly (recommended repetition_penalty=1.2). It is best suited for short to medium-length professional sentences.

Dataset

Trained on the MohamedAshraf701/sentence-corrector dataset, which contains pairs of informal/rude sentences and their polite equivalents.