snxtyle/IndicOCR-MarGuj

IndicOCR MarGuj

LoRA for Marathi and Gujarati line OCR on the recognizer inside bodhan-ai/indic-ocr (weights/ocr, Qwen3.5-0.8B).

The same adapter was trained in two steps. It was first trained on the Marathi real-10k set, then continued for one epoch on 9,200 Gujarati gazette lines. The Gujarati run did not start a new LoRA. Vision weights stay frozen. Base weights are not in this repo.

Training

	Marathi step	Gujarati step
Starting weights	bodhan-ai/indic-ocr weights/ocr	this adapter, after the Marathi step
Lines	9,200 real Marathi lines	9,200 real Gujarati lines
Optimizer steps	1,150	1,150
Batch	1, gradient accumulation 8	1, gradient accumulation 8
Learning rate	5e-5	1e-4
LoRA	r=16, alpha=32, language-model projections only	same adapter, continued

Prompt: Transcribe the text in this image. Write any mathematical expressions in LaTeX, using $...$ for inline math and $$...$$ for display equations.

Images are area-clamped between 256 and 1536 pixels on a side. Generation uses tokenizer EOS id 262146.

Gujarati frozen 100

Same 100 long lines (special_chars_expansion.jsonl) and the same Gujarati metric definitions. PARSeq runs are piecewise. Chitrapathak and IndicOCR runs are whole-line.

Metric	PARSeq base	PARSeq trained	Chitrapathak base	Chitrapathak MarGuj	IndicOCR base	IndicOCR MarGuj
CER ↓	31.57%	27.45%	80.35%	8.81%	1.76%	1.23%
Strict WER ↓	97.30%	99.44%	93.10%	32.25%	8.57%	7.36%
Punctuation F1 ↑	0.00%	36.66%	90.44%	97.30%	97.30%	98.60%
Matra ↑	84.01%	91.87%	0.00%	92.10%	98.53%	98.53%
Conjunct ↑	78.70%	89.51%	0.00%	64.20%	97.22%	97.53%
Numeral ↑	54.60%	82.18%	0.00%	92.53%	96.55%	95.98%
Character accuracy ↑	68.43%	72.55%	19.65%	91.19%	98.24%	98.77%
Exact match ↑	0/100	0/100	0/100	5/100	43/100	46/100

Gujarati 800-line holdout

Short gazette lines, at most 10 words. This is the split held out from Gujarati training. IndicOCR MarGuj: CER 0.014%, exact match 797/800, vowel signs 100%, conjuncts 100%, numerals 100%, punctuation F1 99.98%.

Use

import torch
from peft import PeftModel
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

base_id = "bodhan-ai/indic-ocr"
adapter_id = "snxtyle/IndicOCR-MarGuj"
device = "cuda" if torch.cuda.is_available() else "mps"
dtype = torch.bfloat16

model = Qwen3_5ForConditionalGeneration.from_pretrained(
    base_id, subfolder="weights/ocr", torch_dtype=dtype, attn_implementation="sdpa"
)
model = PeftModel.from_pretrained(model, adapter_id).to(device).eval()
processor = AutoProcessor.from_pretrained(base_id, subfolder="weights/ocr")

prompt = (
    "Transcribe the text in this image. Write any mathematical expressions in LaTeX, "
    "using $...$ for inline math and $$...$$ for display equations."
)
messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": prompt},
    ],
}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], padding=True, return_tensors="pt").to(device)
eos = 262146
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=512, do_sample=False, eos_token_id=eos)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[-1]:], skip_special_tokens=True)[0].strip())