Apply for AI Grants India

What AutoTrain can—and cannot—do

Hugging Face AutoTrain simplifies supervised fine-tuning, but it does not replace decisions about data quality, task design, or evaluation. For Marathi projects, those decisions matter because datasets may be smaller, spelling conventions vary, and text often mixes Devanagari with English, numerals, emojis, and transliterated Marathi.

AutoTrain is a good fit for tasks such as:

• Text classification: sentiment, topic, intent, toxicity, or support-ticket routing.
• Token classification: named-entity recognition, such as people, places, organisations, and schemes.
• Extractive question answering: identifying an answer span in a Marathi passage.
• Causal language-model fine-tuning: adapting an instruction or completion model to a defined Marathi style, provided you have sufficient compute and carefully formatted examples.

For broader model adaptation, first review best practices for fine-tuning LLMs on custom data. AutoTrain is most useful when the objective is narrow, measurable, and supported by representative labelled examples.

1. Define the Marathi use case and label policy

Write the task specification before collecting data. Define what the model should predict, what each label means, and how ambiguous examples will be handled. For example, a Marathi news classifier might distinguish politics, education, health, and agriculture; it should also document whether an article can receive multiple labels.

Create a short annotation guide in Marathi and English. Include examples for:

• Code-mixed Marathi-English sentences.
• Marathi written in Latin transliteration.
• Regional vocabulary and dialect variation.
• Spelling variants, punctuation, and informal chat language.
• Content involving caste, religion, politics, health, or personal data.

Keep a small gold set reviewed by at least two fluent annotators. Use it to measure agreement and to catch label drift before training. Do not assume that a larger scraped corpus is automatically better: duplicated headlines, copied articles, boilerplate, and machine-translated text can make validation scores look stronger than actual performance.

2. Prepare a clean, leakage-resistant dataset

For text classification, use a CSV or JSONL file with one example per row. A typical structure is:

text,label
"हा चित्रपट कुटुंबासोबत पाहता येतो",positive
"सेवेबद्दल ग्राहक समाधानी नाही",negative

Use the column names required by the AutoTrain task configuration, and confirm the current documentation before uploading because supported schemas and interface options can change. Keep labels consistent, without accidental spaces or mixed casing.

For token classification, store tokens and their aligned entity labels. For instruction-style fine-tuning, use a consistent prompt-response format and ensure that the target response—not the prompt—is what the training objective should reward.

Before splitting the data:

• Normalise only what is safe; do not erase meaningful punctuation or Marathi diacritics.
• Detect duplicate and near-duplicate text.
• Record the source, licence, date, and annotation status for every example.
• Remove private or sensitive information unless you have a lawful, documented reason to retain it.
• Separate train, validation, and test data by document or source, not just by random row.

A practical starting split is 80/10/10, but stratify classification labels and preserve rare categories in each split where possible. Keep the test set untouched until model selection is complete.

3. Choose a base model deliberately

Start with a model that already represents Marathi well enough for your task. Compare multilingual encoders such as xlm-roberta-base or bert-base-multilingual-cased with a Marathi-focused or Indian-language checkpoint available on the Hugging Face Hub. Check the model card for language coverage, tokenizer behaviour, licence, training data, and known limitations.

The tokenizer is especially important. Inspect how it segments common Marathi words, inflections, names, and code-mixed text. Excessive fragmentation can increase sequence length and reduce effective context. Run a small baseline before fine-tuning so you know whether improvements come from the model or from data changes.

If your goal is generative regional-language capability rather than classification, compare this workflow with fine-tuning Llama for Indian regional languages. A smaller model with a strong Marathi dataset may be cheaper and easier to deploy than a larger general-purpose model.

4. Create and configure an AutoTrain project

Sign in to Hugging Face, open AutoTrain, create a project, and select the task matching your dataset. Upload the training and validation files, choose the base model, and map the text and label columns carefully. Before starting a paid run, launch a small test job to verify that parsing, tokenisation, labels, and evaluation all work as expected.

Useful starting settings for an encoder classification task are:

• Learning rate: test 2e-5 and 5e-5 rather than assuming one value is best.
• Epochs: begin with 2–4 and use validation results to detect overfitting.
• Batch size: use the largest stable value allowed by available memory; gradient accumulation can help.
• Maximum sequence length: measure Marathi text lengths first instead of choosing an unnecessarily large limit.
• Weight decay and warm-up: enable conservative regularisation where supported.
• Seed: record it and repeat promising runs with more than one seed.

Save the project configuration, dataset version, base-model revision, and run identifier. Reproducibility is essential when comparing experiments or applying for funding.

5. Evaluate Marathi quality, not just one score

Accuracy can hide poor performance on minority labels. Report macro-F1, per-class precision and recall, a confusion matrix, and performance by text type. For imbalanced datasets, macro-F1 is usually more informative than aggregate accuracy.

Build an evaluation slice containing:

• Pure Devanagari Marathi.
• Code-mixed and transliterated text.
• Short messages and long documents.
• Dialectal or informal language.
• New sources not represented in training.

Then perform manual error analysis. Group failures into unclear labels, tokenisation problems, missing vocabulary, source shift, and harmful or unsafe predictions. Ask Marathi-speaking reviewers whether the output is linguistically natural and socially appropriate; numerical metrics cannot answer that question alone.

If you are adapting a generative model, include human review for factuality, instruction following, repetition, and refusal behaviour. Reducing repetitive responses in LLM applications offers useful considerations for that stage.

6. Export, deploy, and monitor

After selecting a checkpoint, publish a model card describing the intended use, Marathi varieties covered, data provenance, licence, metrics, limitations, and known failure cases. Do not expose a private dataset through logs, samples, or public repositories.

For a prototype, use a Hugging Face endpoint or a small API service. For production, measure latency, memory use, throughput, and cost on the actual hardware. Quantisation or distillation may help when the model must run on a low-cost server or mobile device; see this 2026 deployment guide for optimising AI models on mobile devices.

Add monitoring for input distribution shifts, confidence changes, label imbalance, and user complaints. Retrain only after reviewing new examples and confirming that they are correctly labelled. Keep a rollback version and never silently replace a model used in public services.

A practical launch checklist

• Define one measurable Marathi task and a written label policy.
• Verify licences, consent, provenance, and privacy requirements.
• Deduplicate data and split by source where leakage is possible.
• Establish a baseline before fine-tuning.
• Test multilingual and Marathi-focused checkpoints.
• Run a small AutoTrain job before spending on full training.
• Report macro-F1 and slice-level results, not accuracy alone.
• Review errors with fluent Marathi speakers.
• Publish limitations and monitor the deployed model.

Fine-tuning with AutoTrain can shorten the path from a Marathi dataset to a working model, but the durable advantage comes from disciplined data curation and evaluation. Treat AutoTrain as an efficient training interface—not as a substitute for language expertise, responsible sourcing, or production testing.