Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,72 @@
|
|
1 |
---
|
2 |
license: bigscience-bloom-rail-1.0
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: bigscience-bloom-rail-1.0
|
3 |
+
language:
|
4 |
+
- fr
|
5 |
+
- en
|
6 |
+
pipeline_tag: text-classification
|
7 |
---
|
8 |
+
|
9 |
+
|
10 |
+
Bloomz-3b-guardrail
|
11 |
+
---------------------
|
12 |
+
|
13 |
+
We introduce the Bloomz-3b-guardrail model, which is a fine-tuning of the [Bloomz-3b-sft-chat](https://huggingface.co/cmarkea/bloomz-3b-sft-chat) model. This model is designed to detect the toxicity of a text in five modes:
|
14 |
+
|
15 |
+
* Obscene: Content that is offensive, indecent, or morally inappropriate, especially in relation to social norms or standards of decency.
|
16 |
+
* Sexual explicit: Content that presents explicit sexual aspects in a clear and detailed manner.
|
17 |
+
* Identity attack: Content that aims to attack, denigrate, or harass someone based on their identity, especially related to characteristics such as race, gender, sexual orientation, religion, ethnic origin, or other personal aspects.
|
18 |
+
* Insult: Offensive, disrespectful, or hurtful content used to attack or denigrate a person.
|
19 |
+
* Threat: Content that presents a direct threat to an individual.
|
20 |
+
|
21 |
+
Training
|
22 |
+
--------
|
23 |
+
|
24 |
+
The training dataset consists of 500k examples of comments in English and 500k comments in French (translated by Google Translate), each annotated with a toxicity severity gradient. The dataset used is provided by [Jigsaw](https://jigsaw.google.com/) as part of a Kaggle competition : [Jigsaw Unintended Bias in Toxicity Classification](https://www.kaggle.com/competitions/jigsaw-unintended-bias-in-toxicity-classification/data). Since the scores represent severity gradients, regression was preferred using the following loss function:
|
25 |
+
$$loss=l_{\mathrm{obscene}}+l_{\mathrm{sexual\_explicit}}+l_{\mathrm{identity\_attack}}+l_{\mathrm{insult}}+l_{\mathrm{threat}}$$
|
26 |
+
with
|
27 |
+
$$l_i=\frac{1}{\vert\mathcal{O}\vert}\sum_{o\in\mathcal{O}}\vert\mathrm{score}_{i,o}-\sigma(\mathrm{logit}_{i,o})\vert$$
|
28 |
+
Where sigma is the sigmoid function and O represents the set of learning observations.
|
29 |
+
|
30 |
+
Benchmark
|
31 |
+
---------
|
32 |
+
|
33 |
+
As the scores range from 0 to 1, a performance measure such as MAE or RMSE may be challenging to interpret. Therefore, Pearson's inter-correlation was chosen as a measure. Pearson's inter-correlation is a measure ranging from -1 to 1, where 0 represents no correlation, -1 represents perfect negative correlation, and 1 represents perfect positive correlation. The goal is to quantitatively measure the correlation between the model's scores and the scores assigned by judges for 750 comments not seen during training.
|
34 |
+
|
35 |
+
| Model | Language | Obsecene (x100) | Sexual explicit (x100) | Identity attack (x100) | Insult (x100) | Threat (x100) | Mean |
|
36 |
+
|-------------------------------------------------------------------------------|----------|:-----------------------:|-------------------------------|-------------------------------|----------------------|----------------------|------|
|
37 |
+
| [Bloomz-560m-guardrail](https://huggingface.co/cmarkea/bloomz-560m-guardrail) | French | 62 | 73 | 73 | 68 | 61 | 67 |
|
38 |
+
| [Bloomz-560m-guardrail](https://huggingface.co/cmarkea/bloomz-560m-guardrail) | English | 63 | 61 | 63 | 67 | 55 | 62 |
|
39 |
+
| [Bloomz-3b-guardrail](https://huggingface.co/cmarkea/bloomz-3b-guardrail) | French | 72 | 82 | 80 | 78 | 77 | 78 |
|
40 |
+
| [Bloomz-3b-guardrail](https://huggingface.co/cmarkea/bloomz-3b-guardrail) | English | 76 | 78 | 77 | 75 | 79 | 77 |
|
41 |
+
|
42 |
+
With a correlation of approximately 60 for the 560m model and approximately 80 for the 3b model, the output is highly correlated with the judges' scores.
|
43 |
+
|
44 |
+
How to Use Blommz-3b-guardrail
|
45 |
+
--------------------------------
|
46 |
+
|
47 |
+
The following example utilizes the API Pipeline of the Transformers library.
|
48 |
+
|
49 |
+
```python
|
50 |
+
from transformers import pipeline
|
51 |
+
|
52 |
+
guardrail = pipeline("text-classification", "cmarkea/bloomz-3b-guardrail")
|
53 |
+
|
54 |
+
list_text = [...]
|
55 |
+
result = guardrail(
|
56 |
+
list_text,
|
57 |
+
return_all_scores=True, # Crucial for assessing all modalities of toxicity!
|
58 |
+
function_to_apply='sigmoid' # To ensure obtaining a score between 0 and 1!
|
59 |
+
)
|
60 |
+
```
|
61 |
+
|
62 |
+
Citation
|
63 |
+
--------
|
64 |
+
|
65 |
+
```bibtex
|
66 |
+
@online{DeBloomzGuard,
|
67 |
+
AUTHOR = {Cyrile Delestre},
|
68 |
+
URL = {https://huggingface.co/cmarkea/bloomz-3b-guardrail},
|
69 |
+
YEAR = {2023},
|
70 |
+
KEYWORDS = {NLP ; Transformers ; LLM ; Bloomz},
|
71 |
+
}
|
72 |
+
```
|