Apply for AI Grants India

Marathi Good Morning messages are a useful test case for language AI: the outputs are short, culturally specific and easy to evaluate with native speakers. A LoRA adapter can teach a compatible language model a consistent style without updating every parameter, making experimentation more affordable for Indian builders.

This guide explains how to generate Marathi WhatsApp Good Morning templates using LoRA fine-tuning. It focuses on a practical workflow: choosing a base model, building a consented dataset, training an adapter, checking Marathi quality and producing messages that are suitable for WhatsApp Business workflows.

Define the output before training

Start with a clear template specification. Decide whether the model should produce:

• One greeting or several alternatives per prompt.
• Devanagari Marathi only, or Marathi with limited English words.
• A devotional, family-friendly, motivational or professional tone.
• Plain text, emojis, line breaks and optional placeholders such as {name}.
• A maximum length, such as 160–300 characters for easy sharing.

Do not train on random forwarded messages collected without permission. Remove phone numbers, names, personal details, copyrighted poems and content that recipients did not agree to share. A smaller, clean dataset is more useful than a large, noisy archive.

For language-specific quality checks, compare your approach with guidance on fine-tuning AI models for Marathi dialects. If you need a Marathi-capable starting point, also review the open-source Marathi language models guide.

Choose a suitable base model

LoRA is not a complete model; it is a lightweight set of trainable updates attached to a base model. Choose a causal, instruction-tuned model that supports text generation and has reasonable Marathi coverage. Check its licence, context length, tokenizer behaviour and performance on Devanagari before committing to training.

A practical selection process is:

1. Generate 20–30 Marathi prompts with the unmodified model.
2. Check spelling, grammar, gender agreement and code-switching.
3. Measure response length and whether the model follows formatting instructions.
4. Confirm that the model can run within your available GPU or cloud budget.

For a narrow greeting use case, a small model with strong Marathi support may outperform a larger general model that frequently produces Hindi or unnatural translations.

Prepare a supervised training dataset

Use instruction-response examples rather than a flat list of greetings. JSONL is convenient for supervised fine-tuning:

{"messages":[{"role":"user","content":"Create a warm Marathi Good Morning message for a close friend. Use one emoji and keep it under 180 characters."},{"role":"assistant","content":"सुप्रभात मित्रा! तुझ्या दिवसाची सुरुवात आनंदाने होवो आणि प्रत्येक क्षण सुंदर जावो. ☀️"}]}

Create varied examples covering family, colleagues, elders, community groups and festival-neutral daily greetings. Include negative or constraint examples where appropriate: no exaggerated promises, no awkward translations, no excessive emojis and no unexplained English phrases.

Keep a held-out validation set that the adapter never sees during training. Review examples with at least one fluent Marathi speaker, preferably from the intended region and audience. Marathi usage varies across Maharashtra, so document preferences for vocabulary, punctuation and dialect rather than assuming one universal style.

Install the training stack

A common 2026 workflow uses Hugging Face Transformers, Datasets, PEFT and TRL. Install versions compatible with your selected model and CUDA environment:

pip install -U transformers datasets peft trl accelerate bitsandbytes

The exact APIs change, so pin tested versions in requirements.txt. LoRA can be combined with 4-bit quantisation, often called QLoRA, to reduce memory requirements. Quantisation improves accessibility but can affect output quality; compare it against a full-precision baseline on a small evaluation set.

Configure LoRA fine-tuning

The important LoRA settings are the target modules, rank, scaling factor and dropout. For decoder models, attention projections such as q_proj and v_proj are common targets, but the correct modules depend on the architecture.

A conceptual PEFT configuration looks like this:

from peft import LoraConfig

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj", "v_proj"],
    task_type="CAUSAL_LM"
)

Use conservative training settings first. Short greetings can overfit quickly, especially when many examples share the same structure. Begin with a low learning rate, one or two epochs and frequent validation. Save checkpoints so you can compare outputs instead of selecting the last checkpoint automatically.

Your training prompt should tell the model exactly how to respond. For example: “Write one natural Marathi Good Morning greeting in Devanagari. Do not add an explanation. Use at most one emoji.” Consistent formatting makes evaluation easier and reduces unwanted commentary.

Evaluate Marathi quality, not just loss

Training loss cannot tell you whether a greeting sounds natural. Build a review rubric covering:

• Language accuracy: spelling, grammar, sentence flow and Devanagari rendering.
• Meaning preservation: the message should match the requested tone and audience.
• Cultural fit: avoid literal translations, forced idioms and inappropriate religious references.
• Diversity: outputs should not repeat one memorised greeting.
• Safety: reject harassment, stereotypes, medical claims and manipulative language.
• Format compliance: honour length, emoji and placeholder requirements.

Ask native reviewers to score blind outputs from the base model and LoRA model. Track repeated phrases and calculate simple checks such as character length, emoji count and placeholder validity. Keep a regression set containing names, gender-neutral prompts, formal messages and mixed Marathi-English requests.

Generate WhatsApp-ready templates

After evaluation, wrap generation in a controlled function that validates the result before delivery:

def clean_template(text, max_chars=300):
    text = text.strip()
    if len(text) > max_chars:
        text = text[:max_chars].rsplit(" ", 1)[0] + "…"
    return text

Use a low-to-moderate temperature for reliable greetings and generate multiple candidates when variety matters. Apply filters for unsafe content, accidental personal data and malformed placeholders. Human approval is sensible for public campaigns or large broadcast lists.

WhatsApp Business messaging also has platform, consent and template-approval requirements. Do not send unsolicited bulk messages, scrape contacts or present AI-generated text as a personal message without appropriate disclosure. For compliant operational workflows, see the WhatsApp Business Calling API for sales teams and automated VoIP calling for WhatsApp Business in India resources, while remembering that calling and messaging have different policies.

Improve the adapter over time

Log prompts, generated outputs, reviewer decisions and delivery metrics without storing unnecessary recipient data. Sample failures weekly and add corrected examples to a versioned dataset. Retrain only when the evidence justifies it; frequent updates on a tiny dataset can make the model less stable.

Useful production metrics include approval rate, Marathi-language error rate, duplicate-output rate, average length and opt-out or complaint rate. Maintain a rollback path to the base model and label every adapter with its dataset version, licence information and evaluation results.

Frequently asked questions

Is LoRA necessary for a few Marathi greetings?
No. Prompting a Marathi-capable model may be sufficient for a small volume. LoRA becomes useful when you need a consistent tone, format or domain vocabulary at scale.

Should Marathi text be converted to lowercase?
No. Lowercasing is not appropriate for Devanagari in the same way it is for Latin text. Preserve script, punctuation and meaningful formatting.

Can I train on forwarded WhatsApp messages?
Only if you have the right to use them and have removed personal information. Forwarded content often contains duplicates, attribution issues and private data.

Can the adapter generate images or audio?
No. A text LoRA adapter generates text. Separate image or voice systems are needed for multimedia greetings, with their own consent and safety checks.