snxtyle/chitrapathak-marathi

Chitrapathak Marathi v7 LoRA

LoRA adapter for krutrim-ai-labs/Chitrapathak-2, fine-tuned for Marathi line OCR on real Maharashtra government resolutions.

This repository contains the adapter only (about 228 MB). Load it on top of Chitrapathak-2. The base weights are not included.

The adapter is a non-commercial research derivative of Chitrapathak-2 and is distributed under the Krutrim Community License Agreement Version 1.0. Attribution: Krutrim.

Evaluation

Same frozen 100 Marathi line images for every model (marathi_ocr_eval_v2). This set was not used to train this adapter.

Overall winner: IndicOCR v7 LoRA 🏆

🏆 marks the best score on that row. Lower is better for CER, WER, and boundary errors. Higher is better for accuracy, F1, and exact match. Tied winners each get a trophy.

Metric	PARSeq Base	PARSeq V3	PARSeq All Real D	Chitrapathak Base	Chitrapathak old LoRA	Chitrapathak v7 LoRA (this model)	IndicOCR v7 LoRA
CER ↓	28.04%	26.70%	24.56%	6.63%	3.90%	3.60%	3.52% 🏆
Content-no-WS CER ↓	21.76%	16.15%	15.56%	3.65%	3.27%	3.18% 🏆	3.22%
Strict WER ↓	96.40%	98.93%	96.49%	37.20%	13.83%	11.78%	10.52% 🏆
Normalized WER ↓	96.40%	98.93%	96.49%	37.20%	13.83%	11.78%	10.52% 🏆
Content-aware WER ↓	56.57%	46.45%	46.45%	7.21%	5.94%	5.84% 🏆	6.04%
Lexical WER ↓	91.57%	58.17%	55.52%	5.52%	3.86% 🏆	4.11%	4.66%
Punctuation F1 ↑	0.00%	76.80%	76.20%	97.45%	97.99% 🏆	97.89%	96.98%
Matra accuracy ↑	86.53%	86.73%	86.82%	96.58% 🏆	96.24%	96.19%	96.58% 🏆
Conjunct accuracy ↑	82.16%	89.67%	84.98%	98.12% 🏆	98.12% 🏆	98.12% 🏆	98.12% 🏆
Numeral accuracy ↑	83.68%	86.56%	86.36%	98.10%	98.86%	98.71%	99.00% 🏆
Character accuracy ↑	71.96%	73.30%	75.44%	93.37%	96.10%	96.40%	96.48% 🏆
Exact match ↑	0/100	0/100	0/100	38/100	52/100	60/100	63/100 🏆
Boundary merges ↓	125	223	217	79	36	25	17 🏆
Boundary splits ↓	6	3	3	0 🏆	0 🏆	0 🏆	1

This adapter leads content-no-whitespace CER and content-aware WER, and ties the best conjunct accuracy and boundary-split count.

Training

• Base: Chitrapathak-2 (Qwen2.5-VL, vision tower frozen)
• Data: 9,200 real Marathi line crops from official government resolutions, at most 10 words per line. 800 lines from other documents were held out of training.
• LoRA rank 32, alpha 64, dropout 0.05, on language-model attention and MLP projections
• 2 epochs, learning rate 1e-4, effective batch 16, bf16, AdamW, cosine schedule, 3% warmup
• Prompt: system “You are a helpful assistant.” plus “Perform OCR on this image and transcribe all visible text exactly as it appears.”
• Image budget during training: max_pixels=300000, min_pixels=3136

Usage

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base = "krutrim-ai-labs/Chitrapathak-2"
adapter = "snxtyle/chitrapathak-marathi-v7-lora"

processor = AutoProcessor.from_pretrained(base, max_pixels=300_000, min_pixels=56 * 56)
model = AutoModelForImageTextToText.from_pretrained(base, dtype=torch.bfloat16)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "line.png"},
            {"type": "text", "text": "Perform OCR on this image and transcribe all visible text exactly as it appears."},
        ],
    },
]