---
license: cc-by-nc-4.0
base_model: facebook/nllb-200-3.3B
language:
  - en
  - it
tags:
  - translation
  - subtitles
  - fine-tuned
  - experimental
  - nllb
  - italian
pipeline_tag: translation
---

# NLLB-200 3.3B — Fine-Tuned EN→IT Subtitles (v3 · Experimental)

> ⚠️ **Experimental / Test Release** — This model has been fine-tuned for research and personal use. It has not been evaluated on standard benchmarks and may produce errors, especially outside the subtitle domain. Use at your own risk.

## Model Description

This is a fine-tuned version of [facebook/nllb-200-3.3B](https://huggingface.co/facebook/nllb-200-3.3B), optimised for **English → Italian** translation of cinematic and TV subtitles, with emphasis on colloquial language, slang, and natural dialogue.

The fine-tuning was performed in three successive rounds (continual learning) without modifying the model architecture.

## Training Details

| Stage | Dataset | Pairs | Best eval_loss |
|---|---|---|---|
| **v1** | Subtitle pairs EN-IT | 576 k | 1.1629 (step 35 500) |
| **v2** | Continual FT — subtitle + TV corpus | 1.99 M | 1.2296 |
| **v3** (Stage 2) | Slang/colloquial — GPT-4 synthetic (21 k) + TV anchors (9 k) | 29.4 k | 1.4008 |

**Optimised for:**
- Cinematic and TV series subtitles
- Colloquial language, slang expressions, natural dialogue
- Register preservation (formal, informal, humorous)

## Usage

```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Chinasky71/nllb-200-3.3b-it-subtitles")
model = AutoModelForSeq2SeqLM.from_pretrained("Chinasky71/nllb-200-3.3b-it-subtitles")

inputs = tokenizer("Hello, how are you?", return_tensors="pt",
                   src_lang="eng_Latn")
output = model.generate(**inputs,
                        forced_bos_token_id=tokenizer.convert_tokens_to_ids("ita_Latn"))
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

## Limitations

- Evaluated only on subtitle-domain text; general-purpose translation quality may be lower than the base model
- v3 slang fine-tuning slightly increased eval_loss on the held-out set — the model trades general accuracy for colloquial fluency
- Not suitable for medical, legal, or safety-critical applications

## Credits & License

This model is derived from **[facebook/nllb-200-3.3B](https://huggingface.co/facebook/nllb-200-3.3B)** developed by **Meta AI** as part of the [No Language Left Behind](https://ai.facebook.com/research/no-language-left-behind/) project.

All credit for the base architecture, pre-training data, and original weights belongs to Meta AI. This fine-tuned derivative is released under the same **[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)** license as the original model.

> **Citation (original model):**
> ```
> @article{nllb2022,
>   title={No Language Left Behind: Scaling Human-Centered Machine Translation},
>   author={{NLLB Team} and Costa-juss\`a, Marta R. and others},
>   journal={arXiv preprint arXiv:2207.04672},
>   year={2022}
> }
> ```