metadata
license: apache-2.0
language:
- en
datasets:
- allenai/prosocial-dialog
GPT-JT-Moderation-6B v1
This model card introduces on a moderation model, a GPT-JT model fine-tuned on allenai/prosocial-dialog.
In chat applications the moderation model runs in tandem with the main chat bot, checking both the user question and the bot answer for any inappropriate content. If needed, the moderation model intervenes overriding the main chat bot’s response and indicating to the user that this request could not be answered.
Examples
An example prompt and its expected result is as follows:
Possible labels:
1. casual
2. needs caution
3. needs intervention
4. possibly needs caution
5. probably needs caution
Input: I want to call in sick
Output: casual
Input: I hate my job and I don't want to go there.
Output: possibly needs caution
Input: It's so hard to deal with this. I hate this job and I want to quit.
Output: possibly needs caution
Input: Something tells me that you're sitting in the wrong section of the Stormfront bus.
Output: needs caution
Input: Yes I am. They're doing terrible things to people and they're getting away with it, so we should do something about these people.
Output:
-----
> needs intervention
Uses
Limitations and Bias
- The model's performance is limited by the quality and representativeness of its training data. We will continue working on this.
- The model may produce false positives or false negatives, leading to unnecessary confusion. We apologize and welcome any feedbacks or comments for that!
Training
Training Data
- allenai/prosocial-dialog.
- A small subset of OpenChat's data to augment
casualqueries.
Training Procedure
- Hardware: 8 x A100 GPUs
- Optimizer: AdamW
- Gradient Accumulations: 1
- Batch: 16 x 4 = 64
- Learning rate: warmup to 1e-5 for 100 steps and then kept constant
Evaluation Results
[TODO]