---
library_name: transformers
license: other
license_name: custom
license_link: LICENSE
model_index:
- name: Llama-speechlmm-1.0-l-ST
base_model:
- meetween/Llama-speechlmm-1.0-l
datasets:
- facebook/covost2
language:
- es
- it
- en
- fr
- de
metrics:
- bleu
pipeline_tag: translation
---
## Model Information
This is the version of [meetween/Llama-speechlmm-1.0-l](https://huggingface.co/meetween/Llama-speechlmm-1.0-l) that was
fine-tuned for Speech-to-Text Translation.
**License:** see [LICENSE](LICENSE)
## Model Architecture
Identical to the base model. The model was obtained by training LoRA on the LLM.
This repository contains the model weights with LoRA merged into the main weights.
## How to Use
Identical to the base model.
## Fine-tuning Data
This model has been fine-tuned on the same EuroParl-ST and CoVoST2 speech translation data ({en, fr, it, de, es} → {en, fr, it, de, es}) from the training data of the base
model.
## Evaluation Results
| DATASET: |
CoVoST2 |
ACL 60/60 |
AVG |
| BLEU |
en-de |
de-en |
es-en |
fr-en |
it-en |
en-fr |
en-de |
| SeamlessM4T |
- |
- |
- |
- |
- |
40.4 |
28.0 |
- |
| SpeechLMM_v1.0_L |
31.1 |
36.2 |
41.1 |
39.0 |
32.5 |
29.1 |
27.6 |
33.8 |
| SpeechLMM_v1.0_L_ST |
33.7 |
36.7 |
41.5 |
39.0 |
32.4 |
29.6 |
28.5 |
34.5 |
## Framework Versions
- Transformers 4.45.0
- Pytorch 2.3.1+cu124.post2
- Datasets 3.2.0
- Tokenizers 0.20.0