See axolotl config

axolotl version: 0.4.0

# Weights and Biases logging config
#wandb_project: llama3-15B-zeroed-aethora
#wandb_entity:
#wandb_watch: all
#wandb_name:
#wandb_log_model:

# MLFlow
mlflow_tracking_uri: http://127.0.0.1:8888
mlflow_experiment_name: Default
hf_mlflow_log_artifacts: false

# Model architecture config
base_model: elinas/Llama-3-15B-Instruct-zeroed
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

# Hugging Face saving config
hub_model_id: elinas/Aethora-15B-v2-QLoRA
hub_strategy: every_save
push_dataset_to_hub:
hf_use_auth_token: true

# Model checkpointing config
output_dir: Aethora-15B-v2-QLoRA
resume_from_checkpoint:
save_steps:
saves_per_epoch: 2
save_safetensors: true
save_total_limit: 2

# Mixed precision training config
bf16: true
fp16: false
tf32: false

# Model loading config
load_in_8bit: false
load_in_4bit: true
strict: false

# Sequence config
sequence_len: 8192
s2_attention: false
sample_packing: true # false
eval_sample_packing: false # true
pad_to_sequence_len: true # false
train_on_inputs: false
group_by_length: false

# QLoRA adapter config
adapter: qlora
lora_model_dir:
lora_r: 64
lora_alpha: 64
lora_dropout: 0.1
lora_fan_in_fan_out:
lora_target_linear:
save_embedding_layers:
peft_layers_to_transform:
peft_use_dora:
peft_use_rslora:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj
lora_modules_to_save:
  - embed_tokens
  - lm_head
# Unfrozen parameters for FFT
unfrozen_parameters:

# Dataset config
# Dataset config
datasets:
  - path: TheSkullery/Aether-Lite-v1.8.1
    type: customllama3
val_set_size: 0.10
evaluation_strategy: steps
eval_steps: 200
evals_per_epoch:
test_datasets:
dataset_prepared_path: last_run_prepared
shuffle_merged_datasets: true

# Training hyperparameters
num_epochs: 4
gradient_accumulation_steps: 8
micro_batch_size: 2
warmup_steps: 10
optimizer: paged_adamw_8bit # adamw_bnb_8bit # paged_adamw_32bit 
lr_scheduler: cosine
learning_rate: 0.0001
loraplus_lr_ratio: 8
loraplus_lr_embedding:
cosine_min_lr_ratio: 0.1
weight_decay: 0.05
max_grad_norm: 1.0
logging_steps: 1

# Model optimization / unsloth ---- INSTALL UNSLOTH
gradient_checkpointing: unsloth
unsloth_cross_entropy_loss: true
unsloth_lora_mlp: true
unsloth_lora_qkv: true
unsloth_lora_o: true

xformers_attention: false
flash_attention: true
sdp_attention: false

# Loss monitoring config
early_stopping_patience: false
loss_watchdog_threshold: 100.0
loss_watchdog_patience: 3

# Debug config
debug: true
seed: 6969 # 42

# DeepSpeed config
#deepspeed: deepspeed_configs/zero3_bf16.json
# deepspeed: deepspeed_configs/zero3_bf16_optimized.json

# Token config
special_tokens:
  bos_token: "<|begin_of_text|>"
  eos_token: "<|end_of_text|>"
  pad_token: "<|end_of_text|>"
tokens:

# Don't mess with this, it's here for accelerate and torchrun
local_rank:

Aethora-15B-v2-QLoRA

This model is a fine-tuned version of elinas/Llama-3-15B-Instruct-zeroed on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.1942

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 2
• eval_batch_size: 2
• seed: 6969
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• 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: 4

Training results

Training Loss	Epoch	Step	Validation Loss
1.7015	0.0067	1	1.7614
1.2247	1.3244	200	1.2017
1.2301	2.6480	400	1.1942

Framework versions

• PEFT 0.10.0
• Transformers 4.40.2
• Pytorch 2.3.0
• Datasets 2.19.1
• Tokenizers 0.19.1