Improve language tag
#2
by
lbourdois
- opened
README.md
CHANGED
|
@@ -1,112 +1,126 @@
|
|
| 1 |
-
---
|
| 2 |
-
library_name: transformers
|
| 3 |
-
license: mit
|
| 4 |
-
datasets:
|
| 5 |
-
- openslr/librispeech_asr
|
| 6 |
-
- slprl/SpokenSwag
|
| 7 |
-
- slprl/sTinyStories
|
| 8 |
-
base_model:
|
| 9 |
-
- Qwen/Qwen2.5-0.5B
|
| 10 |
-
pipeline_tag: audio-to-audio
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
-
|
| 38 |
-
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
```
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
license: mit
|
| 4 |
+
datasets:
|
| 5 |
+
- openslr/librispeech_asr
|
| 6 |
+
- slprl/SpokenSwag
|
| 7 |
+
- slprl/sTinyStories
|
| 8 |
+
base_model:
|
| 9 |
+
- Qwen/Qwen2.5-0.5B
|
| 10 |
+
pipeline_tag: audio-to-audio
|
| 11 |
+
language:
|
| 12 |
+
- zho
|
| 13 |
+
- eng
|
| 14 |
+
- fra
|
| 15 |
+
- spa
|
| 16 |
+
- por
|
| 17 |
+
- deu
|
| 18 |
+
- ita
|
| 19 |
+
- rus
|
| 20 |
+
- jpn
|
| 21 |
+
- kor
|
| 22 |
+
- vie
|
| 23 |
+
- tha
|
| 24 |
+
- ara
|
| 25 |
+
---
|
| 26 |
+
|
| 27 |
+
# Model Card for SLAM
|
| 28 |
+
|
| 29 |
+
This is a Speech Language Model trained for generating speech continuations over discrete [Hubert tokens](https://huggingface.co/slprl/mhubert-base-25hz).
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Model Details
|
| 33 |
+
|
| 34 |
+
### Model Description
|
| 35 |
+
This is a Speech Language Model, introduced in "[_Slamming_: Training a Speech Language Model on One GPU in a Day](https://arxiv.org/abs/2502.15814)", focusing on efficient training.
|
| 36 |
+
It was fine-tuned from [Qwen/Qwen2.5-0.5B](https://huggingface.co/Qwen/Qwen2.5-0.5B) over a vocabulary of 500 speech tokens extracted from
|
| 37 |
+
the 11-th layer of [mhubert-25hz](https://huggingface.co/slprl/mhubert-base-25hz). For a stronger version of the model trained with
|
| 38 |
+
slightly more compute - 2*A100 for 2 days, see [slam_scaled](https://huggingface.co/slprl/slam_scaled).
|
| 39 |
+
|
| 40 |
+
The model was trained by next-token prediction over a subset of LibriSpeech, Libri-Light and a synthetic data
|
| 41 |
+
[sTinyStories](https://huggingface.co/datasets/slprl/sTinyStories). It was then trained with DPO over
|
| 42 |
+
[SpokenSwag](https://huggingface.co/datasets/slprl/SpokenSwag).
|
| 43 |
+
|
| 44 |
+
- **Developed by:** [SLP-RL](https://huggingface.co/slprl)
|
| 45 |
+
- **Model type:** SpeechLM
|
| 46 |
+
- **License:** MIT
|
| 47 |
+
- **Finetuned from model:** [Qwen/Qwen2.5-0.5B](https://huggingface.co/Qwen/Qwen2.5-0.5B)
|
| 48 |
+
|
| 49 |
+
### Model Sources
|
| 50 |
+
|
| 51 |
+
- **Repository:** [https://github.com/slp-rl/slamkit](https://github.com/slp-rl/slamkit)
|
| 52 |
+
- **Paper:** [https://arxiv.org/abs/2502.15814](https://arxiv.org/abs/2502.15814)
|
| 53 |
+
- **Demo:** [https://pages.cs.huji.ac.il/adiyoss-lab/slamming/](https://pages.cs.huji.ac.il/adiyoss-lab/slamming/)
|
| 54 |
+
|
| 55 |
+
## Uses
|
| 56 |
+
This is a base SpeechLM and as such can be used to generate continuations for speech segments, or as base for further tuning. See the _SlamKit_
|
| 57 |
+
[codebase](https://github.com/slp-rl/slamkit) for more details on usage, and checkout the [demo page](https://pages.cs.huji.ac.il/adiyoss-lab/slamming/) for some generation examples
|
| 58 |
+
|
| 59 |
+
### Out-of-Scope Use
|
| 60 |
+
This model was trained on curated speech datasets which contain mainly audio-books and stories, as such the outputs should not be treated as factual in any way.
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
## How to Get Started with the Model
|
| 65 |
+
We refer users to the official repository for full usage explanations - [github](https://github.com/slp-rl/slamkit).
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
## Training Details
|
| 69 |
+
We highly encourage users to read the full [paper](https://arxiv.org/abs/2502.15814), for full training details, a brief overview is provided below.
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
### Training Data
|
| 73 |
+
This model was trained on a subset of [LibriSpeech](https://huggingface.co/datasets/openslr/librispeech_asr) train,
|
| 74 |
+
[Libri-Light](https://ai.meta.com/tools/libri-light/) and the synthetic dataset
|
| 75 |
+
[sTinyStories](https://huggingface.co/datasets/slprl/sTinyStories) for the pre-training phase. It was also trained with DPO on the synthetic
|
| 76 |
+
dataset [SpokenSwag](https://huggingface.co/datasets/slprl/SpokenSwag).
|
| 77 |
+
|
| 78 |
+
### Training Procedure
|
| 79 |
+
This model was trained by next token prediction over several datasets, and then trained with DPO over [SpokenSwag](https://huggingface.co/datasets/slprl/SpokenSwag).
|
| 80 |
+
Please refer to the [paper]() or [code](https://github.com/slp-rl/slamkit) for the full training recipes.
|
| 81 |
+
|
| 82 |
+
#### Preprocessing
|
| 83 |
+
Speech tokens are extracted from the audio using [Hubert-25hz](https://huggingface.co/slprl/mhubert-base-25hz), and quantised using the
|
| 84 |
+
official kmeans released with the model in [textlesslib](https://github.com/facebookresearch/textlesslib/tree/main). Units are de-duplicated.
|
| 85 |
+
We encourage you to explore the official repository for full details - [github](https://github.com/slp-rl/slamkit).
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
## Evaluation
|
| 89 |
+
The paper provides full results, we do give here some results and also refer to the [demo page]() to listen to some samples.
|
| 90 |
+
| Model | Compute (GPU days) | Parameters | sBLIMP ↑ | sStoryCloze ↑ | tStoryCloze ↑ | GenPPL ↓ | Auto-BLEU ↓ |
|
| 91 |
+
|------------------------------------------|--------------------|------------|----------|--------------|--------------|---------|------------|
|
| 92 |
+
| [TWIST-1.3B](https://pages.cs.huji.ac.il/adiyoss-lab/twist/) | 160xV100 | 1B | 57.00 | 52.4 | 70.6 | 131.8 | 3.20 |
|
| 93 |
+
| [TWIST-7B](https://pages.cs.huji.ac.il/adiyoss-lab/twist/) | ? | 7B | 59.00 | 55.3 | 74.1 | 93.7 | 3.06 |
|
| 94 |
+
| [TWIST-13B](https://pages.cs.huji.ac.il/adiyoss-lab/twist/) | ? | 13B | 59.20 | 55.4 | 76.4 | - | - |
|
| 95 |
+
| [Scaled Optimal](https://arxiv.org/abs/2404.00685) | ? | 823M | **61.3** | **56.7** | **78.0** | - | - |
|
| 96 |
+
| [Predicted Optimal]((https://arxiv.org/abs/2404.00685)) | 1xA5000 | 78M | 56.85 | 54.09 | 70.49 | - | - |
|
| 97 |
+
| TWIST-350M (Original recipe) | 1xA5000 | 305M | 51.52 ± .19 | 53.65 ± .57 | 68.80 ± .47 | 259.2 ± 6.7 | 3.26 ± .46 |
|
| 98 |
+
| *Slam (-DPO) (ours)* | 1xA5000 | 358M | *56.45* ± .17 | *55.59* ± .30 | *78.01* ± .27 | *88.3* ± 1.0 | 3.47 ± .17 |
|
| 99 |
+
| **Slam (ours)** | 1xA5000 | 358M | **58.86** ± .20 | **58.04** ± .51 | **82.04** ± .21 | **62.8** ± 4.1 | 3.88 ± .11 |
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
### Compute Infrastructure
|
| 104 |
+
This model was trained as part of ["*Slamming*: Training a Speech Language Model on One GPU in a Day"](https://arxiv.org/abs/2502.15814), focusing on efficient training.
|
| 105 |
+
|
| 106 |
+
#### Hardware
|
| 107 |
+
This model was trained using **only a single Nvidia A5000 GPU**, 16 CPU cores and 24 GB of RAM for **24 hours**.
|
| 108 |
+
|
| 109 |
+
#### Software
|
| 110 |
+
The model was trained using the [*SlamKit*](https://github.com/slp-rl/slamkit) codebase which builds upon 🤗transformers extending it to support
|
| 111 |
+
easy and efficient training of Speech Language Models.
|
| 112 |
+
|
| 113 |
+
## Citation
|
| 114 |
+
|
| 115 |
+
**BibTeX:**
|
| 116 |
+
```
|
| 117 |
+
@misc{maimon2025slamming,
|
| 118 |
+
title={Slamming: Training a Speech Language Model on One GPU in a Day},
|
| 119 |
+
author={Gallil Maimon and Avishai Elmakies and Yossi Adi},
|
| 120 |
+
year={2025},
|
| 121 |
+
eprint={2502.15814},
|
| 122 |
+
archivePrefix={arXiv},
|
| 123 |
+
primaryClass={cs.LG},
|
| 124 |
+
url={https://arxiv.org/abs/2502.15814},
|
| 125 |
+
}
|
| 126 |
```
|