LimaRP-perscengen-v5

LimaRP Persona-Scenario Generator (v5, Alpaca)

A previously unpublished LoRA adapter for Yarn-Llama-2-7B-64k made for internal use. Its primary purpose is generating Persona and Scenario (summary) from LimaRP yaml source data. To some extent it can work with different text types, however.

Prompt format

### Input:
{Your text here}

### Response:
Charactername's Persona: {output goes here}

Replace Charactername with the name of the character you want to infer a Persona for. By default this LoRA looks for the placeholder names <FIRST> and <SECOND> (in this respective order) but it can work with proper names as well.

Example

This image shows what would happen (red box) after adding data in the format shown in the left pane.

In practice the results would be double-checked and manually tweaked to diversify the outputs and adding character quirks, peculiarities or traits that the model couldn't catch.

Known issues

• While the scenario/summary is often remarkably accurate, personas don't show a very high accuracy and can be repetitive.
• Persona and Scenario may exhibit gpt-isms.
• Peculiar character quirks may not be observed by the model.
• The LoRA hasn't been extensively tested with different input formats.
• There are apparently issues with the EOS token getting generated too early. It's suggested to disable it.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.00025
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• total_train_batch_size: 8
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 2

Training results

Framework versions

• Transformers 4.34.0.dev0
• Pytorch 2.0.1+cu118
• Datasets 2.14.5
• Tokenizers 0.14.0