See axolotl config

axolotl version: 0.4.1

base_model: meta-llama/Meta-Llama-3-70B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: afrias5/AlpacaJustOriginalFeedback
    type: alpaca
    field: text

# resume_from_checkpoint: ~/scratch/70B_070424
val_set_size: 0.20
output_dir: ~/scratch/70BThursday6
sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true
adapter: lora
lora_model_dir: ~/scratch/70BThursday5
lora_r: 8
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
# wandb_project: '70BBCheckpoint'
wandb_entity:
wandb_watch:
wandb_run_id: 'z757bx3a'
# wandb_name: 'lora_70B'
wandb_log_model:

gradient_accumulation_steps: 16
micro_batch_size: 2
num_epochs: 6
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.0002
hub_model_id: afrias5/70BTest
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false
save_steps:
save_total_limit: 1
gradient_checkpointing: true
early_stopping_patience:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
s2_attention:
eval_strategy: "epoch"
warmup_steps: 10
eval_sample_packing: False
evals_per_epoch: 1 
eval_table_size:
eval_max_new_tokens: 10
# saves_per_epoch: 1 
debug:
save_strategy:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.0 #prevent overfitting since using small dataset
fsdp:
save_safetensors:
fsdp_config:
special_tokens:
   pad_token: <|end_of_text|>

70BTest

This model is a fine-tuned version of meta-llama/Meta-Llama-3-70B on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.3704

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: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 16
• total_train_batch_size: 64
• total_eval_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: 6

Training results

Training Loss	Epoch	Step	Validation Loss
0.487	1.0	1	1.3882
0.5958	2.0	2	1.3882
0.1585	2.1667	3	1.3543
0.3686	3.0	4	1.3543
0.6433	4.0	5	1.3543
0.227	4.1667	6	1.3704

Framework versions

• PEFT 0.11.1
• Transformers 4.41.1
• Pytorch 2.2.2+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1