jai23023/indicstr12-signboards-preprocessed

INDICSTR12 Signboards — Hindi / Tamil / Marathi / Telugu

Variant: images in this repo have need-based preprocessing applied (per-image low-light / denoise / exposure / contrast / sharpen, only where a lightweight no-reference quality classifier flagged a deficiency; clean images are passed through unchanged). See the raw variant at jai23023/indicstr12-signboards.

Curated subset of the INDICSTR12 signboard OCR dataset (verified-twice split), reformatted for vision-language instruction tuning. Built to fine-tune Gemma 3n E2B on Indic-script signboard reading using Unsloth + QLoRA.

Contents

Split	Rows
train	907
validation	101
total	1008

Language	Images
Hindi	173
Tamil	335
Marathi	202
Telugu	298

Schema

Column	Type	Description
image	Image	RGB photo of a real-world signboard
language	string	One of hindi, tamil, marathi, telugu
prompt	string	Instruction phrasing (4 variants for robustness)
transcription	string	Ground-truth text, space-joined in GT-file order

Usage with Unsloth (Gemma 3n E2B, QLoRA)

from datasets import load_dataset
from unsloth import FastVisionModel
from trl import SFTTrainer, SFTConfig
from unsloth.trainer import UnslothVisionDataCollator

ds = load_dataset("jai23023/indicstr12-signboards-preprocessed")

def to_messages(example):
    return {
        "messages": [
            {"role": "user", "content": [
                {"type": "image", "image": example["image"]},
                {"type": "text",  "text":  example["prompt"]},
            ]},
            {"role": "assistant", "content": [
                {"type": "text", "text": example["transcription"]},
            ]},
        ]
    }

ds = ds.map(to_messages)

model, tokenizer = FastVisionModel.from_pretrained(
    "unsloth/gemma-3n-E2B-it", load_in_4bit=True,
)
model = FastVisionModel.get_peft_model(
    model,
    finetune_vision_layers=True,
    finetune_language_layers=True,
    r=16, lora_alpha=16, lora_dropout=0,
)

trainer = SFTTrainer(
    model=model, tokenizer=tokenizer,
    train_dataset=ds["train"], eval_dataset=ds["validation"],
    data_collator=UnslothVisionDataCollator(model, tokenizer),
    args=SFTConfig(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        num_train_epochs=2,
        learning_rate=2e-4,
        bf16=True, optim="adamw_8bit",
        remove_unused_columns=False,
        dataset_kwargs={"skip_prepare_dataset": True},
    ),
)
trainer.train()

Source & Processing

• Source: INDICSTR12 verified_twice split (manually verified ground truth).
• Each source image came with a _gt.txt file containing polygon coordinates and per-word text. Words were concatenated in GT-file order with spaces to form the full-sign transcription.
• Images without GT files were dropped (1 in Tamil: image 49, which has no _gt.txt in the source).
• 90/10 train/validation split (seed 42).

Citation

If you use this dataset, please cite the original INDICSTR12 paper.