This is a fine-tuned version of Gemma 3N specifically optimized for conducting trauma assessments with children from conflict zones. The model is designed to facilitate empathetic, culturally-sensitive conversations and generate professional psychological reports.

Model Description

Base Model: Gemma 3N (unsloth/gemma-3n-E2B-it)
Fine-tuning Method: LoRA (r=16, alpha=16)
Training Focus: Specialized for trauma assessment conversations and report generation
Languages: Multilingual support (Arabic dialects, Ukrainian, English)
Context Length: 2048 tokens

Training Details

Dataset and Training Script: Details about training script and datast curation can be found here: https://github.com/Dahimi/Gemma3n_Finetune_Child_Trauma
Fine-tuning Configuration:
- Batch Size: 1 with gradient accumulation (4 steps)
- Learning Rate: 2e-4 with linear scheduler
- Training Focus: Conversation responses only (loss masked on user inputs)
- LoRA Parameters: r=16, alpha=16
- Context Length: 2048 tokens

Intended Uses

This model is designed for:

Conducting initial trauma assessments with children from conflict zones
Supporting mental health professionals and volunteers
Generating standardized trauma assessment reports
Facilitating multilingual mental health support

Primary Functions

Conversational Assessment:
- Empathetic dialogue with parents/caregivers
- Culturally-appropriate questioning
- Trauma indicator identification
Report Generation:
- Structured professional assessments
- Severity scoring (1-10 scale)
- Risk indicator analysis
- Cultural context documentation

Language Support

The model supports:

Palestinian/Levantine Arabic
Sudanese Arabic
Ukrainian
English

Note: All assessment reports are generated in professional English regardless of conversation language

Limitations & Biases

This model should not replace professional mental health assessment
Should be used as a support tool under professional supervision
May have limitations in understanding extremely specific cultural contexts
Not a substitute for emergency mental health intervention

Training Data

The model was trained on a specialized dataset containing:

Conversation Examples: Multi-turn dialogues between parents and AI
Report Generation Examples: Professional assessment report templates

Data format follows Gemma's chat template with alternating roles.

Usage

python

from transformers import TextStreamer
from unsloth import FastModel

model, tokenizer = FastModel.from_pretrained(
    "SoufianeDahimi/child_trauma_gemma_finetune",
    max_seq_length = 2048,
    load_in_4bit = True,
)

def generate_response(prompt):
    messages = [{
        "role": "user",
        "content": [{"type": "text", "text": prompt}]
    }]
    inputs = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt = True,
        return_tensors = "pt",
        tokenize = True,
        return_dict = True,
    ).to("cuda")
    
    return model.generate(
        **inputs,
        max_new_tokens = 512,
        temperature = 0.7,
        top_p = 0.95,
        top_k = 64,
        streamer = TextStreamer(tokenizer, skip_prompt = True),
    )

Output Format

The model generates two types of responses:

Conversational Responses:
- Empathetic dialogue
- Follow-up questions
- Guidance and support
Assessment Reports:
- Parent observations summary
- Trauma indicator analysis
- Severity scoring
- Risk assessment
- Cultural context notes

Ethical Considerations

This model is designed for sensitive mental health contexts and should be used with appropriate care:

Always use under professional supervision
Maintain strict privacy and data protection
Consider cultural and contextual sensitivities
Follow ethical guidelines for mental health assessment

Citations

If you use this model in your research or application, please cite: