base_model: meta-llama/Meta-Llama-3.1-8B # same model you originally used
peft_model: ahmedelgebaly/llama-3.1-8b-squadv2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: ahmedelgebaly/SciQ_Alpaca
    type: alpaca
    split: train

test_datasets:
  - path: ahmedelgebaly/SciQ_Alpaca
    type: alpaca
    split: validation

dataset_prepared_path:
output_dir: ./outputs/qlora-out

adapter: qlora
lora_model_dir:

sequence_len: 2048 # Halves memory usage decreasing from 4096

sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

lora_r: 64 # Increased from 32
lora_alpha: 32 # Increased from 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: llama-3.1-8b-squadv2_SciQ_e1_v2
wandb_entity:
wandb_watch:
wandb_name: llama-3.1-8b-squadv2-v0_SciQ_e1_v2
wandb_log_model:

hub_model_id: ahmedelgebaly/llama-3.1-8b-squadv2_SciQ_e1_v2

gradient_accumulation_steps: 32 # Keeps effective batch size=64 (2x32)
micro_batch_size: 2 # Decrreses from 4
num_epochs: 1
optimizer: paged_adamw_32bit
lr_scheduler: cosine_with_restarts # Updated
learning_rate: 0.0001 # Reduced from 0.0002

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: 100 # Increased from 10
evals_per_epoch: 4
eval_table_size:
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: "<|end_of_text|>"

llama-3.1-8b-squadv2_SciQ_e1_v2

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B on the None dataset. It achieves the following results on the evaluation set:

Loss: 1.5100

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: 42
gradient_accumulation_steps: 32
total_train_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine_with_restarts
lr_scheduler_warmup_steps: 100
num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.8006	0.0598	1	1.8330
1.7825	0.2393	4	1.8315
1.7629	0.4785	8	1.8140
1.6663	0.7178	12	1.7312
1.5168	0.9570	16	1.5100

Framework versions

PEFT 0.13.2
Transformers 4.45.2
Pytorch 2.3.1+cu121
Datasets 3.0.1
Tokenizers 0.20.1