GPT-J-Pyg_PPO-6B [GPT-J Pygmalion Dev V8p4 + GPT-J PPO_HH]
GPT-J-Pyg_PPO-6B is an experimental model containing a parameter-wise 40/60 blend (weighted average PPO_HH:Pygmalion) of the weights of ppo_hh_gpt-j and Pygmalion-6b Dev V8p4.
-Intended Merge Value-
As with fine-tuning, merging weights does not add information but transforms it, therefore it is important to consider trade-offs. Pyg_PPO combines ppo_hh_gpt-j and Pygmalion-6b; both technical achievements are blended with the intent to elevate the strengths of both. Datasets of both are linked below to assist in exploratory speculation on which datasets in what quantity and configuration have the largest impact on the usefulness of a model without the expense of fine-tuning. Blend was done in FP32 and output in FP16.
Research purposes only, intended for responsible use. Express a conversation in natural language, and Pyg_PPO will do the thing. 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 format.
Can also be used as a base to merge with other creative, technical, or adventure themed models of the same class (GPT-J & 6b NeoX) and parameter size (6b) to experiment with the morphology of model weights based on the value added by instruct.
Merge tested using KoboldAI with Nucleus Sampling Top-P set to 0.9, Temperature at 0.6, and Repetition Penalty at 1.1; extra samplers disabled.
Core Model: https://huggingface.co/EleutherAI/gpt-j-6B Author: https://www.eleuther.ai/
Model1; 50% ppo_hh_gpt-j: https://huggingface.co/reciprocate/ppo_hh_gpt-j
Author Repo: https://huggingface.co/reciprocate
Related; CarperAI: https://huggingface.co/CarperAI
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
Model2; 50% Pygmalion-6b: https://huggingface.co/PygmalionAI/pygmalion-6b
Author Repo: https://huggingface.co/PygmalionAI
Weight merge Script credit to Concedo: https://huggingface.co/concedo
Model's card template credit to Digitous: https://huggingface.co/digitous/GPT-R
- Downloads last month