limarp-miqu-1-70b-qlora

<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.4.0

base_model: models/miqu-1-70b-sf
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer
is_llama_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: train-all-max-alpaca-llama.jsonl
    type: completion
dataset_prepared_path:
val_set_size: 0.0
output_dir: ./volume/limarp-70b-qlora

adapter: qlora
lora_model_dir:

sequence_len: 16384
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: 70b-lora
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0001

train_on_inputs: true
group_by_length: false
bf16: true
fp16: false
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
eval_steps:
eval_table_size:
save_steps:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

</details><br>

limarp-miqu-1-70b-qlora

Experimental limarp qlora trained at 16384 ctx length (greater than size of the longest limarp sample when tokenized via llama's tokenizer) on the fixed dequantized miqu-1-70b model by 152334H.

I wasn't particularly happy with the results I got when I tried applying the lora at varying weights to the miqu-1-70b model. It's possible that this is related to the fact that the model was dequantized from Q5_K_M GGUF, or perhaps due to it already being an instruct-tuned model.

However, I decided to go ahead and release this in case someone else finds a use for it. Provided as-is and YMMV.

Model description

The intended prompt format is the Alpaca instruction format of LimaRP v3:

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

User's Persona: {user character description}

Scenario: {what happens in the story}

Play the role of Character. Taking the above information into consideration, 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.)

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).

Intended uses & limitations

The model will show biases similar to those observed in niche roleplaying forums on the Internet, besides those exhibited by the base model.

Training and evaluation data

For more details about LimaRP, see the dataset page.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 2

Framework versions

• PEFT 0.7.2.dev0
• Transformers 4.37.0
• Pytorch 2.1.2+cu118
• Datasets 2.16.1
• Tokenizers 0.15.0