Generic Gemma 4	MentalBridge
"Here are 5 tips: 1) Exercise 2) Sleep..."	"That feeling of wanting to disappear before facing them — I understand why every step toward home feels impossible right now."
Gives American crisis numbers	Routes to iCall 9152987821 and Vandrevala Foundation
Western therapy language	Understands IIT pressure, CA failures, arranged marriage
English only	Responds naturally in Tamil, Hindi, Telugu

Training Details

Base model: unsloth/gemma-4-E2B-it
Method: LoRA via Unsloth
Platform: Kaggle 2x Tesla T4
Dataset: 192 examples (Kaggle mental health intents + 14 custom Indian scenarios)
Languages: English, Tamil, Hindi
Final training loss: 0.843
Epochs: 3 · Steps: 72 · LoRA rank: r=16

Indian Scenarios Covered

NEET / CA / UPSC exam failure and family shame
Arranged marriage pressure
Job loss and financial shame
Software engineer burnout
Tamil and Hindi language support
Crisis escalation to Indian helplines only

How to Use

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "vishnuprasadgs/mentalbridge-gemma4-E2B-finetuned",
    max_seq_length=512,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content": "I failed my CA exam and can't face my parents."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(input_ids=inputs, max_new_tokens=150, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Ethical Commitments

Never diagnoses any condition
Always recommends professional help
Crisis responses route to Indian helplines only — iCall 9152987821, Vandrevala 1860-2662-345
Runs 100% on-device — zero data leaves the phone

Application

Powers the MentalBridge app — offline mental health companion for India. GitHub: https://github.com/vishnu-prasad-g-s/mentalbridge

Uploaded finetuned model

Developed by: vishnuprasadgs
License: apache-2.0
Finetuned from model : unsloth/gemma-4-E2B-it

This gemma4 model was trained 2x faster with Unsloth and Huggingface's TRL library.