RPMax is a series of models that are trained on a diverse set of curated creative writing and RP datasets with a focus on variety and deduplication. This model is designed to be highly creative and non-repetitive by making sure no two entries in the dataset have repeated characters or situations, which makes sure the model does not latch on to a certain personality and be capable of understanding and acting appropriately to any characters or situations.
Early tests by users mentioned that these models does not feel like any other RP models, having a different style and generally doesn't feel in-bred.
Llama-3.1-70B-ArliAI-RPMax-v1.2 is a variant made from the Meta-Llama-3.1-70B-Instruct model.
Let us know what you think of the model! The 8B and 12B versions of RPMax had great feedback from users, so we expect this 70B version to one of the best RP models.
v1.2 update completely removes non-creative/RP examples in the dataset and is also an incremental improvement of the RPMax dataset which dedups the dataset even more and better filtering to cutout irrelevant description text that came from card sharing sites.
Specs
Context Length: 128K
Parameters: 70B
Training Details
Sequence Length: 4096
Training Duration: Approximately 5 days on 2x3090Ti
Epochs: 1 epoch training for minimized repetition sickness
LORA: 64-rank 128-alpha, resulting in ~2% trainable weights
Learning Rate: 0.00001
Gradient accumulation: Very low 32 for better learning.
<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are [character]. You have a personality of [personality description]. [Describe scenario]<|eot_id|><|start_header_id|>user<|end_header_id|>
{{ user_message_1 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{{ model_answer_1 }}<|eot_id|><|start_header_id|>user<|end_header_id|>
{{ user_message_2 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
