llama-wp.yml

base_model: meta-llama/Llama-3.1-8B
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer
seed: 1

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: /home/venkat/Desktop/intergroup-nfl/data/train_wp.json
    ds_type: json
    type: alpaca
    split: train

test_datasets:
  - path: /home/venkat/Desktop/intergroup-nfl/data/eval_wp.json
    ds_type: json
    type: alpaca
    split: train

load_best_model_at_end: True
dataset_prepared_path: /home/venkat/Desktop/intergroup-nfl/data/
output_dir: /home/venkat/Desktop/intergroup-nfl/models/llama31-8b-wp-1/

sequence_len: 2560
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false

gradient_accumulation_steps: 1
micro_batch_size: 2
eval_batch_size: 2
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 1e-5
cosine_min_lr_ratio: 0.1
weight_decay: 0.1
warmup_steps: 10

train_on_inputs: false
group_by_length: false
bf16: auto
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
logging_steps: 10
flash_attention: true

eval_steps: 0.1
save_steps: 0.1
save_total_limit: 3
early_stopping_patience: 3
special_tokens:
   pad_token: <|end_of_text|>

wandb_project: intergroup-bias
wandb_name: llama31-8b-fft-wp-2

# deepspeed: deepspeed_configs/zero3_bf16.json