---
base_model: facebook/seamless-m4t-v2-large
library_name: peft
tags:
- base_model:adapter:facebook/seamless-m4t-v2-large
- lora
- transformers
---

# LEO Seamless-M4T v2 Large Roverplastik Adapter

This repository contains the LoRA adapter for **LEO**, a domain-adapted multilingual translation model for Roverplastik technical content.

The adapter is fine-tuned from:

```text
facebook/seamless-m4t-v2-large
```

It is intended for technical translations among Italian, English, French, and Spanish, with a focus on window systems, installation manuals, insulation, monoblocchi, cassonetti, airtightness, and related building-envelope terminology.

## Model Details

- **Base model:** `facebook/seamless-m4t-v2-large`
- **Adapter type:** PEFT LoRA
- **Training framework:** PyTorch Lightning + QLoRA
- **Precision:** bf16 mixed precision, 4-bit base model loading
- **Hardware used:** NVIDIA RTX 4090
- **Released adapter path:** `leo_hf_release`

## Training Checkpoint

The exported adapter was selected from the best validation checkpoint:

```text
seamless-m4t-v2-large-finetuned-epoch=05-val_loss=0.54.ckpt
```

The benchmark was run directly from this `.ckpt` before export. The Hugging Face adapter export is used only for release and deployment.

## Evaluation

Benchmark command:

```bash
python scripts/leo.py benchmark
```

Evaluation set:

- `data/gold/test_set.csv`
- 1005 translation samples
- directions among Italian, English, French, and Spanish

Global results:

| Metric | Base Seamless-M4T | LEO adapter | Absolute delta |
|--------|-------------------|-------------|----------------|
| BLEU | 0.3173 | 0.5806 | +0.2633 |
| chrF | 0.5320 | 0.7382 | +0.2061 |
| METEOR | 0.4749 | 0.7255 | +0.2505 |

Regression check:

```text
152 / 1005 samples regressed by chrF = 15.1%
```

The model shows a strong global improvement over the base model, especially on technical terminology and domain-specific phrasing. Remaining regressions are mostly concentrated in selected non-Italian cross-language directions and in minor function-word differences.

## Usage

The adapter is designed to be loaded with PEFT on top of `facebook/seamless-m4t-v2-large`.

Example with the project CLI:

```bash
python scripts/leo.py infer \
  --src-lang eng_Latn \
  --tgt-lang ita_Latn \
  --text "This window profile ensures excellent air tightness."
```

Expected smoke-test output:

```text
Questo profilo di finestra garantisce un'eccellente tenuta all'aria.
```

The public Space uses the `ADAPTER_PATH` environment variable to point to this adapter repository.

## Intended Use

Use this model for assisted translation and review of Roverplastik-style technical material:

- installation manuals
- technical datasheets
- product catalog text
- building-envelope documentation
- multilingual internal validation workflows

The model is not a substitute for human review in legal, certified, contractual, or safety-critical documentation.

## Limitations

- The benchmark set includes synthetic and curated examples; final production quality should still be checked by domain experts.
- Some regressions remain versus the base model, especially in selected cross-language directions.
- The model may occasionally alter small function words or produce acceptable paraphrases that score lower under character-level metrics.
- The model is optimized for the Roverplastik/building-envelope domain, not for broad general-purpose translation.

## Framework Versions

- PEFT 0.18.0
- Transformers 4.57.3