Shaimaz/gemma3_4B_LoRA_Emotion_Toxicity_v1

from huggingface_hub import HfApi

readme_text = """

🌿 Gemma-3-4B LoRA — Emotion + Toxicity Core (v1)

Author: Shaima Z. Alzubi
Base model: google/gemma-3-4b-it
Technique: LoRA (Low-Rank Adaptation)
Trained Cores: Core-1 (Emotions) + Core-2 (Toxicity)

---

🧠 Overview

This model adapts Gemma-3-4B-IT using a dual-core dataset:

Core	Domain	Objective	Size
🩵 Core-1	Emotion understanding	Detect human emotional states from text	49K
💢 Core-2	Toxicity moderation	Detect & rephrase toxic or disrespectful content	65K

The LoRA improves emotional recognition and safe conversational tone, making the base Gemma more empathetic and socially aware.

---

💬 Example Usage

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from peft import PeftModel

base = "google/gemma-3-4b-it"
adapter = "Shaimaz/gemma3_4B_LoRA_Emotion_Toxicity_v1"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base)
model = PeftModel.from_pretrained(model, adapter)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)

prompt = "Detect the emotion or toxicity in this text and respond kindly:\\n\\nYou are so annoying!"
print(pipe(prompt, max_new_tokens=60)[0]['generated_text'])