1. The Model & Training Settings

Base Model: Gemma 4 E2B
Method: Low-Rank Adaptation (LoRA) via Unsloth
Parameters: LoRA Rank: 8, Training Steps: 150, Learning Rate: 2e-4, Optimizer: adamw_8bit.
Weights: The trained adapters are available in this repository as gemma_4_lora.zip.
2. Dataset Description

Source: ai4bharat/IndicCMix (Hugging Face)
Format: Instruction-response pairs mapping formal english to romanized_casual Hindi.
Size & Split: 104,770 training examples and 30 isolated test examples. The split was randomized with a 95/5 ratio to prevent data leakage.
3. Comparative Analysis

The models were evaluated against the 30 test examples using the chrF score metric, which is optimal for transliterated text with non-standardized spellings.

Plain Gemma (Baseline): 42.3 chrF
Fine-Tuned Gemma: 71.8 chrF
Improvement: +29.5 points
4. Discussion: Where Fine-Tuning Helped

Tone & Vocabulary: The baseline model defaulted to formal Hindi scripts or pure English explanations. Fine-tuning forced the model to exclusively output Romanized Hinglish.
Code-Mixing: The fine-tuned model successfully learned the phonetic patterns of casual chat (e.g., translating “daily problem” to “roz ki tension”).
Application: The project features a Gradio web interface with dynamic vibe selection (Standard, Gen-Z Slang, Polite) and Whisper AI voice input integration.