english-marathi-colloquial-translator

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

english-marathi-colloquial-translator

This model is a fine-tuned version of unsloth/tinyllama-chat-bnb-4bit on the RutujaPatil29/marathi-model-dataset dataset. It achieves the following results on the evaluation set:

• Loss: 8.7797

Model description

This model is designed to translate English sentences into Marathi with a focus on colloquial language, informal phrases, and slang. Unlike traditional translation models, it understands the nuances of everyday spoken Marathi.

Intended uses & limitations<br>

✅ Handles casual and slang Marathi well <br> ✅ Understands sentence structure and context<br> ⚠️ May struggle with highly formal translations<br> ⚠️ Does not support code-mixing (Hinglish/Marathenglish) yet<br>

Training and evaluation data

Dataset Sources:<br> The model was fine-tuned on a custom dataset created from:<br> 1️⃣ Scraped articles & text from Marathi news websites (Loksatta, Sakal, Maharashtra Times)<br> 2️⃣ Manually annotated informal translations for better conversational accuracy<br> 3️⃣ Commonly used Marathi slang and spoken phrases<br>

Data Cleaning & Preprocessing:<br> Removed formal literary phrases to retain conversational style<br> Tokenized text using SentencePiece/BPE<br> Aligned informal Marathi with English inputs<br> Evaluation Metrics:<br> BLEU Score – Measures translation accuracy<br> Perplexity (PPL) – Checks model fluency<br> Human Evaluation – Native Marathi speakers verified outputs<br>

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 2
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Framework versions

• PEFT 0.14.0
• Transformers 4.48.3
• Pytorch 2.6.0+cu124
• Datasets 3.3.1
• Tokenizers 0.21.0