🌸 Marathi Mitra — v2
Fine-tuned Phi-3 Mini for Marathi vocabulary learning using QLoRA (SFT).
Model Description
| Property | Value |
|---|---|
| Base Model | microsoft/Phi-3-mini-4k-instruct |
| Fine-tuning Method | QLoRA (Supervised Fine-Tuning) |
| LoRA Rank | r=32, alpha=64 |
| Training Examples | 250 Marathi vocabulary items |
| Best Experiment | Exp4 (lr=2e-4, epochs=25, r=32) |
| Train/Eval Split | 80/20 |
| Training Hardware | Google Colab T4 GPU |
Performance
| Words | Score |
|---|---|
| Seen words (training set) | 100.0% |
| Unseen words (generalisation) | 78.8% |
| Overall | 89.4% |
Evaluation criteria:
- Field presence (40%) — all 5 sections present
- Exact match (60%) — correct Marathi word + pronunciation
What It Does
Given an English word, generates a Marathi lesson:
- Downloads last month
- 53
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for ninadp/marathi-mitra-phi3-v2
Base model
microsoft/Phi-3-mini-4k-instruct