PaddleOCR-VL-Tachiwin

TachiwinOCR

for the Indigenous Languages of Mexico

32 bits full precision

This is a PaddleOCR-VL Finetune specialized in the 68 indigenous languages of Mexico and their diverse character and glyph repertoire making a world first in tech access and linguistic rights

Inference

You can perform inference using the PaddleOCR pipeline or the transformers library.

Option A: Using PaddleOCR (Easy Pipeline)

from paddleocr import PaddleOCRVL

# Load the fine-tuned model
pipeline = PaddleOCRVL(
    vl_rec_model_name="PaddleOCR-VL-0.9B",
    vl_rec_model_dir=path_to_tachiwin_downloaded_model,
)

# Predict on an image
output = pipeline.predict("test.png")

for res in output:
    res.print()
    res.save_to_json(save_path="output")
    res.save_to_markdown(save_path="output")

Option B: Using Transformers (Advanced Control)

from PIL import Image
import torch
from transformers import AutoModelForCausalLM, AutoProcessor

# ---- Settings ----
model_path = "tachiwin/PaddleOCR-VL-Tachiwin"
image_path = "test.png"
# ------------------

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

image = Image.open(image_path).convert("RGB")

model = AutoModelForCausalLM.from_pretrained(
    model_path, trust_remote_code=True, torch_dtype=torch.bfloat16
).to(DEVICE).eval()
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

messages = [
    {"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": "OCR:"},
    ]}
]

inputs = processor.apply_chat_template(
    messages, 
    tokenize=True, 
    add_generation_prompt=True, 	
    return_dict=True,
    return_tensors="pt"
).to(DEVICE)

outputs = model.generate(**inputs, max_new_tokens=1024, min_new_tokens=1)
generated_text = processor.batch_decode(outputs, skip_special_tokens=True)[0]
print(generated_text)

---

📊 Benchmark Results

Tachiwin-OCR was evaluated against the base PaddleOCR-VL model using a diverse subset of Indigenous language samples. The fine-tuning results demonstrate significant improvements in both character and word recognition accuracy.

Summary Metrics

Detailed Comparison (Sample)

A subset of the evaluation results across different languages, where tonal languages are the most improved by this fine-tuning:

Key Findings

• High Accuracy Gains: In many tonal languages like Tepehuán (stp) and Mazatec (maa), the fine-tuning process reduced the error rate from significant levels to nearly zero or double digits.
• Robustness: The model shows high resilience against synthetic distortions implemented during the data generation phase.
• Word-Level Performance: The relative reduction in Word Error Rate (WER) highlights the model's improved capability in contextualizing character sequences specific to these language families.

Tachiwin (from Totonac - "Language") is dedicated to bridging the digital divide for indigenous languages of Mexico through AI technology.

• Developed by: Tachiwin
• License: apache-2.0
• Finetuned from model : PaddlePaddle/PaddleOCR-VL