Update README.md

README.md

@@ -1,10 +1,86 @@
-#TODO card. Mix of (GPT-J-6B-Janeway + PPO_HH_GPT-J) + Pygmalion-6b
+---
+language: en
+license: apache-2.0
+commercial: 'no'
+inference: false
+---
+# GPT-J 6B - PPO_Pygway Mix
+## Model description
+This is a a merged model, using an averaged weights strategy at a (20:20:60) ratio between the models:
 
-At a ratio of
+- [20%] - KoboldAI/GPT-J-6B-Janeway: https://huggingface.co/KoboldAI/GPT-J-6B-Janeway
+- [20%] - reciprocate/ppo_hh_gpt-j: https://huggingface.co/reciprocate/ppo_hh_gpt-j
+- [60%] - Pygmalion/Pygmalion-6b: https://huggingface.co/Pygmalion/Pygmalion-6b
+ 
+By their respective authors.
 
+**Warning: Pygmalion may generate NSFW or inappropriate content due to being trained on general user logs, and internet archives.**
 
-GPT-J-6B-Janeway - 20%
+### Intended Use:
 
-PPO_HH_GPT-J - 20%
+Research purposes only, intended for responsible use.
+Express a conversation in natural language, and PPO_Pygmalion will pick up on the conversational format.
+Try starting a two line prompt such as:
+```
+Bot: "Hello, how are you?"
+You: "I am doing just fine, thank you."
+```
+Or any other topic, and the model will carry on in this back and forth style.
 
-Pygmalion-6b - 60%
+## Information:
+For more details, check out the related source models, especially [Pygmalion-6b](https://huggingface.co/Pygmalion/Pygmalion-6b) for more information on how to utilize the chat bot formatting expected.
+
+In a similar manner to fine-tuning, merging weights does not add information but transforms it, therefore it is important to consider trade-offs.
+PPO_Pygway combines `ppo_hh_gpt-j`, `Janeway-6b` and `Pygmalion-6b`; all three models were blended in a two step process using the a simple weighted parameter method
+```
+(X*A + Y*B)
+```
+With X & Y being the model weighs, and A/B being how strongly they are represented within the final value. 
+The intent of this is to elevate the end-model by borrowing the strongly represented aspects out of each base model.
+
+Blend was done in FP32 and output saved in FP16 for reduced storage needs.
+
+
+## Limitations and biases
+Based on known problems with NLP technology, potential relevant factors include bias (gender, profession, race and religion).
+**Warning: This model has a very strong NSFW bias!**
+
+### License
+GPT-J-6b is licensed by EleutherAI under the apache-2.0 license. All Rights Reserved.
+
+### BibTeX entry and citation info
+```
+@misc{gpt-j,
+  author = {Wang, Ben and Komatsuzaki, Aran},
+  title = {{GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model}},
+  howpublished = {\url{https://github.com/kingoflolz/mesh-transformer-jax}},
+  year = 2021,
+  month = May
+}
+```
+
+### Credits To:
+
+Models involved:
+- https://huggingface.co/EleutherAI/gpt-j-6B
+- https://huggingface.co/Pygmalion/Pygmalion-6b
+- https://huggingface.co/reciprocate/ppo_hh_gpt-j
+- https://huggingface.co/KoboldAI/GPT-J-6B-Janeway
+
+Average weights merging Script credit to Concedo:
+- https://huggingface.co/concedo
+
+### Related datasets and articles:
+
+PPO_HH-GPT-J-6b's Dataset is a variant of the Helpful Harmless assistant themed
+dataset and Proximal Policy Optimization, specific datasets
+used are unknown; listed repo datasets include:
+- https://huggingface.co/datasets/reciprocate/summarize_eval_ilql
+- https://huggingface.co/datasets/reciprocate/hh_eval_ilql
+
+PPO explained:
+- https://paperswithcode.com/method/ppo
+
+Potential HH-type datasets utilized:
+- https://huggingface.co/HuggingFaceH4
+- https://huggingface.co/datasets/Anthropic/hh-rlhf