See axolotl config

axolotl version: 0.4.1

adapter: lora
base_model: MLP-KTLim/llama-3-Korean-Bllossom-8B
bf16: auto
chat_template: llama3
dataset_prepared_path: null
datasets:
- data_files:
  - ff0278723740ee77_train_data.json
  ds_type: json
  format: custom
  path: /workspace/input_data/ff0278723740ee77_train_data.json
  type:
    field_input: input
    field_instruction: instruction
    field_output: output
    format: '{instruction} {input}'
    no_input_format: '{instruction}'
    system_format: '{system}'
    system_prompt: ''
debug: null
deepspeed: null
early_stopping_patience: null
eval_max_new_tokens: 128
eval_table_size: null
evals_per_epoch: 4
flash_attention: false
fp16: true
fsdp: null
fsdp_config: null
gradient_accumulation_steps: 4
gradient_checkpointing: true
group_by_length: false
hub_model_id: leixa/bb84a5d5-4732-4e64-883e-c7860dd12459
hub_repo: null
hub_strategy: checkpoint
hub_token: null
learning_rate: 0.0001
load_in_4bit: false
load_in_8bit: false
local_rank: 0
logging_steps: 3
lora_alpha: 128
lora_dropout: 0.1
lora_fan_in_fan_out: true
lora_model_dir: null
lora_r: 64
lora_target_linear: true
lr_scheduler: cosine
max_memory:
  0: 72GB
max_steps: 100
micro_batch_size: 8
mlflow_experiment_name: /tmp/ff0278723740ee77_train_data.json
model_type: AutoModelForCausalLM
num_epochs: 3
optimizer: adamw_bnb_8bit
output_dir: miner_id_24
pad_to_sequence_len: true
resume_from_checkpoint: null
s2_attention: false
sample_packing: false
saves_per_epoch: 4
sequence_len: 1024
strict: false
tf32: false
tokenizer_type: AutoTokenizer
train_on_inputs: false
trust_remote_code: true
val_set_size: 0.05
wandb_entity: leixa-personal
wandb_mode: online
wandb_name: bb84a5d5-4732-4e64-883e-c7860dd12459
wandb_project: Gradients-On-Demand
wandb_run: your_name
wandb_runid: bb84a5d5-4732-4e64-883e-c7860dd12459
warmup_steps: 10
weight_decay: 0.01
xformers_attention: null

bb84a5d5-4732-4e64-883e-c7860dd12459

This model is a fine-tuned version of MLP-KTLim/llama-3-Korean-Bllossom-8B on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.3363

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: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• training_steps: 100

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0021	1	3.2399
1.1132	0.0190	9	0.8033
0.3899	0.0381	18	0.4629
0.4748	0.0571	27	0.4546
0.3983	0.0762	36	0.4109
0.3716	0.0952	45	0.3802
0.4076	0.1142	54	0.3659
0.3365	0.1333	63	0.3567
0.277	0.1523	72	0.3462
0.3885	0.1713	81	0.3425
0.3626	0.1904	90	0.3368
0.3685	0.2094	99	0.3363

Framework versions

• PEFT 0.13.2
• Transformers 4.46.0
• Pytorch 2.5.0+cu124
• Datasets 3.0.1
• Tokenizers 0.20.1