update model card

Author: jackclarksf

model_card.md

@@ -1,6 +1,6 @@
 # GPT-2 model card
 
-Last updated: August 2019
+Last updated: November 2019
 
 Inspired by [Model Cards for Model Reporting (Mitchell et al.)](https://arxiv.org/abs/1810.03993), we’re providing some accompanying information about the GPT-2 family of models we're releasing.
 
@@ -10,12 +10,16 @@ This model was developed by researchers at OpenAI to help us understand how the
 
 ### Model date
 
-Spring 2019, trained on data that cuts off at the end of 2017.
+February 2019, trained on data that cuts off at the end of 2017.
 
 ### Model type
 
 Language model
 
+### Model version
+
+1.5 billion parameters: the fourth and largest GPT-2 version. We have also released 124 million, 355 million, and 774 million parameter models.
+
 ### Paper or other resource for more information
 [Blog post](https://openai.com/blog/better-language-models/) and [paper](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf)
 
@@ -42,7 +46,7 @@ Here are some secondary use cases we believe are likely:
 
 Because large-scale language models like GPT-2 do not distinguish fact from fiction, we don’t support use-cases that require the generated text to be true.
 
-Additionally, language models like GPT-2 reflect the biases inherent to the systems they were trained on, so we do not recommend that they be deployed into systems that interact with humans unless the deployers first carry out a study of biases relevant to the intended use-case.
+Additionally, language models like GPT-2 reflect the biases inherent to the systems they were trained on, so we do not recommend that they be deployed into systems that interact with humans unless the deployers first carry out a study of biases relevant to the intended use-case. We found no statistically significant difference in gender, race, and religious bias probes between 774M and 1.5B, implying all versions of GPT-2 should be approached with similar levels of caution around use cases that are sensitive to biases around human attributes.
 
 ## Evaluation Data
 
@@ -60,5 +64,6 @@ The motivation behind WebText was to create an Internet-scale, heterogeneous dat
 
 Because GPT-2 is an internet-scale language model, it’s currently difficult to know what disciplined testing procedures can be applied to it to fully understand its capabilities and how the data it is trained on influences its vast range of outputs. We recommend researchers investigate these aspects of the model and share their results.
 
-Additionally, as indicated in our discussion of issues relating to potential misuse of the model, it remains unclear what the long-term dynamics are of detecting outputs from these models. Developing better approaches to detection today will give us greater intuitions when thinking about future models and could help us understand ahead of time if detection methods will eventually become ineffective.
+Additionally, as indicated in our discussion of issues relating to potential misuse of the model, it remains unclear what the long-term dynamics are of detecting outputs from these models. We conducted [in-house automated ML-based detection research](https://github.com/openai/gpt-2-output-dataset/tree/master/detector) using simple classifiers, zero shot, and fine-tuning methods. Our fine-tuned detector model reached accuracy levels of approximately 95%. However, no one detection method is a panacea; automated ML-based detection, human detection, human-machine teaming, and metadata-based detection are all methods that can be combined for more confident classification. Developing better approaches to detection today will give us greater intuitions when thinking about future models and could help us understand ahead of time if detection methods will eventually become ineffective.
+