# 📚 Bilingual EPUB Translator (English → English/Vietnamese)

Translate any English `.epub` book into a beautifully formatted bilingual English-Vietnamese edition. The LLM runs **directly in Python** via HuggingFace Transformers — no Ollama or external server needed.

The model downloads once from HuggingFace Hub and is cached at `~/.cache/huggingface` for all future runs.

## Prerequisites

- **Python 3.10+** (conda env recommended)
- **256 GB RAM** (for qwen2.5:72b in INT8 quantization)

## Installation

```bash
conda activate env03
cd /home/x79/book_translator
pip install -r requirements.txt
```

> **Note:** The first run will download the model weights (~72 GB for INT8). This is a one-time download cached at `~/.cache/huggingface`.

## Usage

### Basic (uses defaults)

```bash
python translate_epub.py "Mona___s_Eyes___40_Thomas_Schlesser__Hildegarde_Serle___40_translator__41___41_.epub"
```

This produces `Mona___s_Eyes_..._bilingual.epub` in the same directory.

### Specify output path

```bash
python translate_epub.py input.epub -o my_bilingual_book.epub
```

### Use a different / smaller model

```bash
# Smaller model for faster inference
python translate_epub.py input.epub --model Qwen/Qwen2.5-7B-Instruct

# Another model entirely
python translate_epub.py input.epub --model meta-llama/Llama-3-8B-Instruct
```

### Disable INT8 quantization (uses more RAM)

```bash
python translate_epub.py input.epub --no-8bit
```

## All Options

| Flag | Default | Description |
|------|---------|-------------|
| `input` | *(required)* | Path to the English `.epub` file |
| `-o, --output` | `<input>_bilingual.epub` | Output file path |
| `--model` | `Qwen/Qwen2.5-72B-Instruct` | HuggingFace model ID |
| `--chunk-size` | `5` | Paragraphs per inference call |
| `--max-new-tokens` | `8192` | Max generated tokens per chunk |
| `--no-8bit` | off | Disable INT8 quantization |
| `--no-resume` | off | Ignore cached progress, re-translate all |
| `--clear-cache` | — | Delete cached checkpoints and exit |

## Checkpoint / Resume

Since CPU inference on a 72B model can take **many hours**, the script saves each chapter's translated HTML to a `.bilingual_cache_*` directory next to your input file.

If the script crashes or you `Ctrl+C`:

```bash
# Just re-run the same command — it picks up where it left off
python translate_epub.py input.epub
```

To force a fresh start:

```bash
python translate_epub.py input.epub --no-resume
# or delete the cache entirely:
python translate_epub.py input.epub --clear-cache
```

## Output Format

The bilingual EPUB contains, for every paragraph:

1. **English** (bold) — the original text
2. *Vietnamese* (italic, red left border) — the translation
3. A **Vocabulary Box** at the end of each chunk with 3–5 difficult words explained

## Troubleshooting

| Problem | Solution |
|---------|----------|
| Out of memory | Use a smaller model: `--model Qwen/Qwen2.5-7B-Instruct` |
| Slow inference | Reduce chunk size: `--chunk-size 3` |
| Model download fails | Check disk space at `~/.cache/huggingface` and internet |
| Poor translation | Try `--no-8bit` for full precision (needs more RAM) |