metadata
base_model: snxtyle/chitrapathak2
library_name: peft
pipeline_tag: image-text-to-text
tags:
- lora
- ocr
- marathi
- gujarati
- chitrapathak
- peft
language:
- mr
- gu
chitrapathak_marguj_mixed
One shared LoRA adapter for Marathi and Gujarati OCR on top of Chitrapathak-2.
What this is
- Base model: Chitrapathak-2
- Adapter type: PEFT LoRA (r=32, α=64)
- Languages: Marathi + Gujarati in a single adapter
- Not language routing and not stacked Marathi-v7 + guj_v1
Load
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor
base = "YOUR_BASE_OR_LOCAL_CHITRAPATHAK2"
adapter = "snxtyle/chitrapathak_marguj_mixed"
model = AutoModelForImageTextToText.from_pretrained(base)
model = PeftModel.from_pretrained(model, adapter)
processor = AutoProcessor.from_pretrained(base)
Frozen-100 (vs original base)
| Language | Exact match | CER | Script accuracy |
|---|---|---|---|
| Marathi | 44/100 (base 26/100) | 4.10% (base 5.56%) | 95% |
| Gujarati | 11/100 (base 0/100) | 10.61% (base 82.73%) | 100% |
Source training checkpoint: bilingual reviewed MPS v2 step-414.