metadata

license: apache-2.0

LimaRP-Llama2-7B-v3 (Alpaca, experimental, 4-bit LoRA adapter)

This is an experimental version of LimaRP for Llama2, using a somewhat updated dataset (1800 training samples) and a 2-pass training procedure. The first pass includes unsupervised tuning on 2800 stories within 4k tokens length and the second is LimaRP.

For more details about LimaRP, see the model page for the previously released version. Most details written there apply for this version as well.

Prompt format

Same as before. It uses the extended Alpaca format, with ### Input: immediately preceding user inputs and ### Response: immediately preceding model outputs. While Alpaca wasn't originally intended for multi-turn responses, in practice this is not a problem; the format follows a pattern already used by other models.

### Instruction:
Character's Persona: {bot character description}

User's Persona: {user character description}

Scenario: {what happens in the story}

Play the role of Character. You must engage in a roleplaying chat with User below this line. Do not write dialogues and narration for User. Character should respond with messages of medium length.

### Input:
User: {utterance}

### Response:
Character: {utterance}

### Input
User: {utterance}

### Response:
Character: {utterance}

(etc.)

Other notes

Replace all the text in curly braces (curly braces included) with your own text.
User and Character should be replaced with appropriate names.

Training procedure

Axolotl was used for training. The model has been trained as a 4-bit LoRA adapter. It's so large because a LoRA rank of 256 was used. It's suggested to merge it to the base Llama2-7B model.

Training hyperparameters

For both passes these settings were used:

learning_rate: 0.0002
lr_scheduler_type: constant
lora_r: 256
lora_alpha: 16
lora_dropout: 0.1
lora_target_linear: True
num_epochs: 1
bf16: True
tf32: True
load_in_4bit: True
adapter: qlora
micro_batch_size: 2
gradient_accumulation_steps: 1
optimizer: adamw_torch

In the second pass, the lora_model_dir option was used to load and train the adapter previously trained on a stories dataset.