metadata

license: apache-2.0

LimaRP-Mistral-7B-v0.1 (Alpaca, 8-bit LoRA adapter)

This is a version of LimaRP for Mistral-7B-v0.1 with about 1900 training samples up to 9k tokens length

For more details about LimaRP, see the model page for the previously released v2 version for Llama-2. Most details written there apply for this version as well. Generally speaking, LimaRP is a longform-oriented, novel-style roleplaying chat model intended to replicate the experience of 1-on-1 roleplay on Internet forums. Short-form, IRC/Discord-style RP (aka "Markdown format") is not supported yet. The model does not include instruction tuning, only manually picked and slightly edited RP conversations with persona and scenario data.

Known issues

Despite performing a few finetuning attempts, including one that followed almost the same procedure as in previous releases, Mistral-7B-v0.1 appears to have strange repetition issues.
Even though benchmarks tell a different story, in practice the model doesn't feel smarter during roleplay than Llama-2-13B.

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.

### Input:
User: {utterance}

### Response:
Character: {utterance}

### Input
User: {utterance}

### Response:
Character: {utterance}

(etc.)

You should:

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

Message length control

Inspired by the previously named "Roleplay" preset in SillyTavern, with this version of LimaRP it is possible to append a length modifier to the response instruction sequence, like this:

### Input
User: {utterance}

### Response: (length = medium)
Character: {utterance}

This has an immediately noticeable effect on bot responses. The lengths using during training are: micro, tiny, short, medium, long, massive, huge, enormous, humongous, unlimited. The recommended starting length is medium. Keep in mind that the AI can ramble or impersonate the user with very long messages.

The length control effect is reproducible, but the messages will not necessarily follow lengths very precisely, rather follow certain ranges on average, as seen in this table with data from tests made with one reply at the beginning of the conversation:

Response length control appears to work well also deep into the conversation. By omitting the modifier, the model will choose the most appropriate response length (although it might not necessarily be what the user desires).

Suggested settings

You can follow these instruction format settings in SillyTavern. Replace tiny with your desired response length:

Text generation settings

Mistral-7B-v0.1 appears to have repetition issues. A low temperature combined with a relatively high repetition penalty and low penalty range (about as long as the prior 2 messages) appears to help:

TFS = 0.90~0.95
Temperature = 0.50~0.55
Repetition penalty = ~1.15
Repetition penalty range = ~512
top-k = 0 (disabled)
top-p = 1 (disabled)

Training procedure

Axolotl was used for training on 2x NVidia A40 GPUs.

The A40 GPUs have been graciously provided by Arc Compute.

The model has been trained as an 8-bit LoRA adapter, and it's so large because a LoRA rank of 256 was also used. The reasoning was that this might have helped the model internalize any newly acquired information, making the training process closer to a full finetune. It's suggested to merge the adapter to the base Mistral-7B-v0.1 model.

Training hyperparameters

learning_rate: 0.0005
lr_scheduler_type: cosine
num_epochs: 2
sequence_len: 9000
lora_r: 256
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: True
bf16: True
fp16: false
tf32: True
load_in_8bit: True
adapter: lora
micro_batch_size: 2
gradient_accumulation_steps: 32
warmup_steps: 2
optimizer: adamw_torch

For the second pass, the lora_model_dir option was used to continue finetuning on the LoRA adapter obtained from the first pass.

Using 2 GPUs, the effective global batch size would have been 128.