Upload 3 files

README.md

@@ -1,10 +1,39 @@
 ---
-title: Modi Project
-emoji: 🦀
-colorFrom: green
-colorTo: yellow
-sdk: static
+title: ModiTrans
+emoji: 📜
+colorFrom: yellow
+colorTo: red
+sdk: gradio
+sdk_version: 5.0.0
+app_file: app.py
 pinned: false
+license: mit
+short_description: Transliterate historic Modi script images to Devanagari
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# ModiTrans — Modi Script to Devanagari
+
+Gradio demo for [`historyHulk/ModiTrans-12B-Gemma-Teacher`](https://huggingface.co/historyHulk/ModiTrans-12B-Gemma-Teacher),
+a LoRA adapter on `google/gemma-3-12b-it` that transliterates scanned images
+of historic Modi script into modern Devanagari text.
+
+## Setup notes before deploying
+
+1. **Accept the gated model license.** Visit the
+   [model page](https://huggingface.co/historyHulk/ModiTrans-12B-Gemma-Teacher)
+   while logged in and accept the access conditions. You'll also need to
+   accept the license for the base model
+   [`google/gemma-3-12b-it`](https://huggingface.co/google/gemma-3-12b-it).
+2. **Add an `HF_TOKEN` secret** to this Space (Settings → Variables and
+   secrets) using a token from an account that has been granted access to
+   both gated repos.
+3. **Hardware:** this Space is configured for **ZeroGPU** (`@spaces.GPU` in
+   `app.py`). ZeroGPU requires a PRO or verified Hugging Face account. If you
+   don't have ZeroGPU access, remove the `spaces` import/decorator in
+   `app.py` and instead select a dedicated GPU hardware tier (T4 or A10G) in
+   the Space settings.
+
+## Files
+
+- `app.py` — Gradio app + inference logic
+- `requirements.txt` — Python dependencies

app.py

@@ -0,0 +1,115 @@
+import spaces
+import torch
+import gradio as gr
+from PIL import Image
+from transformers import AutoProcessor, AutoModelForImageTextToText
+from peft import PeftModel
+
+BASE_MODEL_ID = "google/gemma-3-12b-it"
+ADAPTER_ID = "historyHulk/ModiTrans-12B-Gemma-Teacher"
+PROMPT = "Translitrate the following Modi script to Devnagri script."
+MAX_NEW_TOKENS = 350
+
+# Model + processor are loaded once at startup and kept on CPU.
+# They are moved to GPU inside the @spaces.GPU-decorated function,
+# which is how ZeroGPU spaces work (GPU only attached per-call).
+print("Loading processor...")
+processor = AutoProcessor.from_pretrained(BASE_MODEL_ID)
+
+print("Loading base model...")
+base_model = AutoModelForImageTextToText.from_pretrained(
+    BASE_MODEL_ID,
+    torch_dtype=torch.bfloat16,
+)
+
+print("Loading LoRA adapter...")
+model = PeftModel.from_pretrained(
+    base_model,
+    ADAPTER_ID,
+    torch_dtype=torch.bfloat16,
+)
+model.eval()
+
+
+@spaces.GPU
+def transliterate(image: Image.Image) -> str:
+    if image is None:
+        return "Please upload an image of Modi script."
+
+    device = "cuda"
+    model.to(device)
+
+    image = image.convert("RGB").resize((1024, 512))
+
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "image": image},
+                {"type": "text", "text": PROMPT},
+            ],
+        },
+    ]
+
+    inputs = processor.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        tokenize=True,
+        return_dict=True,
+        return_tensors="pt",
+    ).to(device, dtype=torch.bfloat16)
+
+    input_len = inputs["input_ids"].shape[-1]
+
+    with torch.no_grad():
+        output_ids = model.generate(
+            **inputs,
+            max_new_tokens=MAX_NEW_TOKENS,
+            do_sample=False,
+        )
+
+    generated = output_ids[0][input_len:]
+    text = processor.decode(generated, skip_special_tokens=True)
+    return text.strip()
+
+
+with gr.Blocks(title="ModiTrans: Modi Script to Devanagari") as demo:
+    gr.Markdown(
+        """
+        # ModiTrans — Modi Script to Devanagari Transliteration
+
+        Upload a scanned image of historic **Modi script** text and this model
+        will transliterate it into modern **Devanagari** script.
+
+        Uses [`historyHulk/ModiTrans-12B-Gemma-Teacher`](https://huggingface.co/historyHulk/ModiTrans-12B-Gemma-Teacher),
+        a LoRA adapter on `google/gemma-3-12b-it`, from the paper
+        *"Historic Scripts to Modern Vision: A Novel Dataset and A VLM Framework
+        for Transliteration of Modi Script to Devanagari"* (ICDAR 2025).
+
+        > This is a gated model — the Space owner's `HF_TOKEN` must have accepted
+        > access on the model page for inference to work.
+        """
+    )
+
+    with gr.Row():
+        with gr.Column():
+            image_input = gr.Image(type="pil", label="Modi Script Image")
+            run_btn = gr.Button("Transliterate", variant="primary")
+        with gr.Column():
+            output_text = gr.Textbox(
+                label="Devanagari Transliteration", lines=8
+            )
+
+    run_btn.click(fn=transliterate, inputs=image_input, outputs=output_text)
+    image_input.change(fn=transliterate, inputs=image_input, outputs=output_text)
+
+    gr.Markdown(
+        """
+        ---
+        **Citation:** Kausadikar, H., Kale, T., Susladkar, O., Mittal, S.
+        *Historic Scripts to Modern Vision.* ICDAR 2025 (Springer LNCS).
+        """
+    )
+
+if __name__ == "__main__":
+    demo.launch()

requirements-1.txt

@@ -0,0 +1,8 @@
+spaces
+torch
+torchvision
+transformers>=4.50.0
+peft
+accelerate
+pillow
+gradio