Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,48 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
+
datasets:
|
4 |
+
- 0-hero/lj_speech_with_spectogram_conversations
|
5 |
+
- 0-hero/lj_speech_with_spectogram
|
6 |
---
|
7 |
+
## Explanation
|
8 |
+
A small experiment insipred by the [Mistral playing DOOM experiment](https://github.com/umuthopeyildirim/DOOM-Mistral/tree/main) from the Mistral Hackathon
|
9 |
+
|
10 |
+
**How it works?**
|
11 |
+
```
|
12 |
+
Audio -> Waveform Visualization -> Waveform ASCII Art -> Finetune Mistral on ASCII Art to predict text from ASCII Art
|
13 |
+
```
|
14 |
+
|
15 |
+
**Quick video explanation**
|
16 |
+
<video width="640" controls src="https://cdn-uploads.huggingface.co/production/uploads/6382255fcae34727b9cc149e/nCB8Qu8QwDbJAKcq9IzPE.mp4"></video>
|
17 |
+
|
18 |
+
## Models & Results
|
19 |
+
Mistral 7B 0.2 finetunes on ascii art. **As seen in the results, experiment didn't amount to much**
|
20 |
+
- 2 epochs - [0-hero/mistral-speech-to-text-preview](https://huggingface.co/0-hero/mistral-speech-to-text-preview/) (this) - Loss pretty much flattened after this epoch
|
21 |
+
- ```
|
22 |
+
Metrics:
|
23 |
+
rouge-1: {'r': 0.12919024091165357, 'p': 0.1163312036605547, 'f': 0.11315199212991178}
|
24 |
+
rouge-2: {'r': 0.013705453572242508, 'p': 0.0137500428446463, 'f': 0.012676757505648992}
|
25 |
+
rouge-l: {'r': 0.11261286554140228, 'p': 0.09921920076529338, 'f': 0.09705621471622536}
|
26 |
+
length_correlation: 0.014470676120233311
|
27 |
+
avg_actual_length: 16.59
|
28 |
+
avg_pred_length: 21.46
|
29 |
+
exact_match_accuracy: 0.0
|
30 |
+
```
|
31 |
+
- 4 epochs - [0-hero/mistral-speech-to-text](https://huggingface.co/0-hero/mistral-speech-to-text/)
|
32 |
+
- ```
|
33 |
+
Metrics:
|
34 |
+
rouge-1: {'r': 0.11869828051815862, 'p': 0.11697319273190071, 'f': 0.11154343875398197}
|
35 |
+
rouge-2: {'r': 0.008572925612399297, 'p': 0.009040061245943597, 'f': 0.008369604666309954}
|
36 |
+
rouge-l: {'r': 0.10780857719316121, 'p': 0.10373665666448233, 'f': 0.09985384905943501}
|
37 |
+
length_correlation: -0.1500200314034927
|
38 |
+
avg_actual_length: 16.59
|
39 |
+
avg_pred_length: 18.32
|
40 |
+
exact_match_accuracy: 0.0
|
41 |
+
```
|
42 |
+
|
43 |
+
## Datasets
|
44 |
+
[lj_speech](https://huggingface.co/datasets/lj_speech) dataset used to convert audio waveforms into ASCII Art
|
45 |
+
- [0-hero/lj_speech_with_spectogram_conversations](https://huggingface.co/datasets/0-hero/lj_speech_with_spectogram_conversations) - ShareGPT style finetuning dataset with train, test split
|
46 |
+
- [0-hero/lj_speech_with_spectogram](https://huggingface.co/datasets/0-hero/lj_speech_with_spectogram) - Raw dataset with ASCII Art
|
47 |
+
|
48 |
+
|