This repository contains the LoRA adapter for LEO, a domain-adapted multilingual translation model for Roverplastik technical content.
The adapter is fine-tuned from:
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:
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:
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:
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.