# Allow cli options to override these settings.
strict: false

# Base model settings.
base_model: meta-llama/Meta-Llama-3-8B-Instruct
tokenizer_config: meta-llama/Meta-Llama-3-8B-Instruct
model_type: AutoModelForCausalLM

# Wandb settings
wandb_entity: collinear
wandb_project: template-training
wandb_name: l3smi-sft-qlora-r64

# Output settings
save_safetensors: true
hub_model_id: fozziethebeat/l3bgi-sft-qlora-r64
dataset_prepared_path: data/l3bgi-sft-qlora-r64
output_dir: models/l3bgi-sft-qlora-r64

# Data format settings
chat_template: llama3
datasets:
  - path: fozziethebeat/alpaca_messages_2k_test
    split: train
    type: chat_template
    chat_template: llama3
    field_messages: messages 
    message_field_role: role
    message_field_content: content
test_datasets:
  - path: fozziethebeat/alpaca_messages_2k_test
    split: test
    type: chat_template
    chat_template: llama3
    field_messages: messages 
    message_field_role: role
    message_field_content: content

# Data packing settings
sequence_len: 512
train_on_inputs: false
pad_to_sequence_len: true
group_by_length: false
sample_packing: false
eval_sample_packing: false

# Adapter settings
adapter: qlora
lora_model_dir:
load_in_8bit: false
load_in_4bit: true
lora_r: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

# Computation Format settings
bf16: true
fp16:
tf32: false

# Trainer settings
gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 1e-5
loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

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

warmup_steps: 10
eval_table_size:
eval_max_new_tokens: 128
evals_per_epoch: 4
saves_per_epoch: 1
debug:
weight_decay: 0.01
special_tokens:
  pad_token: <|end_of_text|>
deepspeed:
fsdp:

Visualize in Weights & Biases

l3bgi-sft-qlora-r64

This model is a fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct on the None dataset. It achieves the following results on the evaluation set:

Loss: 1.0220

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: 1e-05
train_batch_size: 2
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 2
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

Training results

Training Loss	Epoch	Step	Validation Loss
1.0859	0.0022	1	1.3374
0.9847	0.2497	111	1.1122
1.203	0.4994	222	1.0451
1.3916	0.7492	333	1.0307
0.7893	0.9989	444	1.0251
1.0244	1.2486	555	1.0228
0.6814	1.4983	666	1.0221
0.9408	1.7480	777	1.0224
1.0832	1.9978	888	1.0220

Framework versions

PEFT 0.11.1
Transformers 4.43.0.dev0
Pytorch 2.3.1+cu121
Datasets 2.19.1
Tokenizers 0.19.1