---
base_model: unsloth/Qwen3.5-0.8B
tags:
  - lora
  - vision
  - ocr
  - historical-documents
  - handwritten-text-recognition
  - unsloth
license: apache-2.0
language:
- en
- es
datasets:
- private
---

# Uploaded  model

- **Developed by:** Ak137
- **License:** apache-2.0
- **Finetuned from model :** unsloth/Qwen3.5-0.8B

This qwen3_5 model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth)

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)


# 🏛️ Historical Spanish OCR LoRA

LoRA adapter for **Qwen3.5-0.8B** fine-tuned on historical Spanish manuscript images.

## Results

| Split | CER |
|---|---|
| Validation (baseline) | 0.1414 |
| Validation (fine-tuned) | 0.0559 |
| **Test (fine-tuned)** | **0.0309** |

## Training configuration

| Param | Value |
|---|---|
| Base model | `Qwen3.5-0.8B` |
| LoRA r / α / dropout | 16 / 32 / 0.0 |
| Learning rate | 1e-4 (cosine schedule) |
| Epochs | 5 |
| Effective batch size | 24 |
| Max image dim | 2048 px |
| Data augmentation | True |
| Train / Val / Test | ~2000 / 500 / 500 samples |

## Text normalisation applied

1. Long-s `ſ` → `s`
2. `ç`/`Ç` → `z`/`Z`
3. Nasal tilde abbreviations: `ã`→`an`, `õ`→`on`, `ẽ`→`en`, `ũ`→`un`, `ĩ`→`in`
4. D-with-stroke `đ` → `de`
5. Stress accents stripped (except `ñ`)
6. Multiple spaces collapsed

## Inference

```python
from unsloth import FastVisionModel
from PIL import Image

model, tokenizer = FastVisionModel.from_pretrained("Ak137/qwen3.5-0.8B-spanish-ocr-lora", load_in_4bit=False)
FastVisionModel.for_inference(model)

image = Image.open("manuscript.jpg")
w, h = image.size
if max(w, h) > 2048:
    scale = 2048 / max(w, h)
    image = image.resize((int(w*scale), int(h*scale)))

messages = [{{"role": "user", "content": [
    {{"type": "text", "text": "Transcribe the text in this historical Spanish manuscript image."}},
    {{"type": "image"}}
]}}]
input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
inputs = tokenizer(image, input_text, add_special_tokens=False, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```