AuraIndustries/Aura-MoE-2x4B-v2

Aura-MoE-2x4B-v2

Introduction

Aura-MoE-2x4B-v2 is a state of the art dedicated roleplaying model designed to fulfill your every desire.

The finetunes used in this merge saw several hundreds of millions of tokens of instruction data. The merge was then healed on 150 million tokens of roleplaying data. A Kahneman-Tversky Optimization was applied to the healed model to give it a unique output style.

By the numbers, this should be a direct improvement over Aura-MoE-2x4B

Developed by Aura Industries, with contributions from Anthracite Org

Model Details

• Model Name: Aura-MoE-2x4B-v2
• Base Model: IntervitensInc/Llama-3.1-Minitron-4B-Width-Base-chatml
• Model Type: Chat Completions
• Prompt Format: ChatML
• License: Apache-2.0
• Language: English
• Max Context: 8,192+ tokens

License

This model is licensed under the Apache 2.0 License.

Quantizations

Static GGUF

Imatrix GGUF

Open LLM Leaderboard Evaluation Results

Coming soon...

Metric	Value
Avg.	N/A
IFEval (0-Shot)	N/A
BBH (3-Shot)	N/A
MATH Lvl 5 (4-Shot)	N/A
GPQA (0-shot)	N/A
MuSR (0-shot)	N/A
MMLU-PRO (5-shot)	N/A

Training Configuration

Click here for Mergekit and Axolotl configs

MoE Merge

base_model: FourOhFour/Zenith_4B
gate_mode: random
dtype: bfloat16
experts_per_token: 1
experts:
  - source_model: FourOhFour/Luxe_4B
  - source_model: FourOhFour/Zenith_4B

SFT

base_model: jeiku/MoEv2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: FourOhFour/RP_Phase
    type: chat_template
    chat_template: chatml
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: jeiku/Writing
    type: completion
    field: text

chat_template: chatml

shuffle_merged_datasets: true
dataset_prepared_path:
val_set_size: 0.01
output_dir: ./output/out

hub_model_id: jeiku/Aura-MoEv2
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true

sequence_len: 8192
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len:

wandb_project: Aura-MoEv2
wandb_entity:
wandb_watch:
wandb_name: Aura-MoEv2
wandb_log_model:

gradient_accumulation_steps: 16
micro_batch_size: 2
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00005

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

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

warmup_steps: 10
evals_per_epoch: 2
eval_table_size:
eval_max_new_tokens:
saves_per_epoch: 1
debug:
deepspeed: 
weight_decay: 0.05
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|finetune_right_pad_id|>

KTO

base_model: jeiku/Aura-MoEv2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

hub_model_id: jeiku/moekto
hub_strategy: "all_checkpoints"
push_dataset_to_hub:
hf_use_auth_token: true

chat_template: chatml

rl: kto
rl_beta: 0.2
kto_desirable_weight: 0.2

datasets:
  - path: anthracite-core/full-opus-chosen-hermes-rejected-kto-v1
    type: chatml.argilla

shuffle_merged_datasets: true
val_set_size: 0.0
output_dir: ./outputs/out

sequence_len: 8192
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: false

wandb_project: moekto
wandb_entity:
wandb_watch:
wandb_name: moekto
wandb_log_model:

gradient_accumulation_steps: 16
micro_batch_size: 2
num_epochs: 2
max_steps: 500

optimizer: adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00001
weight_decay: 0.05

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

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

warmup_steps: 10
evals_per_epoch: 2
eval_table_size:
eval_max_new_tokens: 
saves_per_epoch: 1

debug:
deepspeed: 
fsdp:
fsdp_config:
fsdp:
fsdp_config:

special_tokens:
  pad_token: <|finetune_right_pad_id|>