See axolotl config

axolotl version: 0.4.1

base_model: meta-llama/Meta-Llama-3-8B
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
  - path: ahmedelgebaly/SQuad_2_Alpaca
    type: alpaca
    split: train
    percentage: 0.1 # small replay buffer to avoid forgetting

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

dataset_prepared_path:
output_dir: ./outputs/qlora-out-mixed

adapter: qlora
lora_model_dir:

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj
#lora_target_linear: true  ##Removed for Explicit Control after adding modules

lora_fan_in_fan_out:

wandb_project: llama-3.1-8b-squadv2_SciQ_e1_v3
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_v3

gradient_accumulation_steps: 4
micro_batch_size: 4
num_epochs: 1
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0001 ##Reducing from 0.0002 to 0.0001
lr_scheduler_warmup_ratio: 0.1 ##Added for better convergence very low for SciQ (only ~11k samples

train_on_inputs: false
group_by_length: true
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
logging_steps: 1
xformers_attention:
flash_attention: true

evals_per_epoch: 4
saves_per_epoch: 1
log_eval_metric: true

weight_decay: 0.0

special_tokens:
  pad_token: "<|end_of_text|>"

llama-3.1-8b-squadv2_SciQ_e1_v3

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

• Loss: 0.8983

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: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.5496	0.0029	1	1.8420
0.9073	0.2504	85	0.9289
0.9158	0.5007	170	0.9048
0.8754	0.7511	255	0.8983

Framework versions

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