base_model: meta-llama/Llama-3.2-3B-Instruct
# Automatically upload checkpoint and final model to HF
hub_model_id: smohammadi/bf16-llama-3B # username/custom_model_name

load_in_8bit: false
load_in_4bit: false
strict: false


    #liger_rope: true
    #liger_rms_norm: true
    #liger_glu_activation: true
    #liger_layer_norm: true
    #
    #liger_fused_linear_cross_entropy: true

datasets:
  - path: yahma/alpaca-cleaned
    type: alpaca
    split: train[:95%]
output_dir: ./outputs/bf16-train_on_inputs/
dataset_prepared_path: ./outputs/ds_prepared_new_token
  #sample_packing: true

sequence_len: 8192
flash_attention: true
  #flex_attention: true
  #flex_attn_compile_kwargs:
  # dynamic: false
  # mode: max-autotune-no-cudagraphs

aosiubdoaisdb:
   activation_dtype: int8
   weight_dtype: int4
   group_size: 32

wandb_project: qat_v2
wandb_entity:
wandb_watch:
wandb_name: bf16-train-on-inputs
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 16
num_epochs: 1
optimizer: adamw_torch_fused
train_on_inputs: true

  #cosine_constant_lr_ratio: 0
  #cosine_min_lr_ratio: 1.0
lr_scheduler: constant
learning_rate: 2e-5
save_only_model: true
bf16: true

resume_from_checkpoint:
logging_steps: 1
include_tkps: true
evals_per_epoch: 1
saves_per_epoch: 1

  #warmup_ratio: 0.1
weight_decay: 0.0

fsdp_config:
  fsdp_version: 2
  fsdp_offload_params: false
  fsdp_cpu_ram_efficient_loading: False
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD
  fsdp_reshard_after_forward: true
  fsdp_activation_checkpointing: true

special_tokens:
  pad_token: <|finetune_right_pad_id|>

bf16-llama-3B

This model is a fine-tuned version of meta-llama/Llama-3.2-3B-Instruct on the yahma/alpaca-cleaned dataset.

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: 2e-05
train_batch_size: 16
eval_batch_size: 16
seed: 42
distributed_type: multi-GPU
num_devices: 4
total_train_batch_size: 64
total_eval_batch_size: 64
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: constant
lr_scheduler_warmup_steps: 23
training_steps: 769

Training results

Framework versions

Transformers 4.55.3
Pytorch 2.7.1+cu126
Datasets 4.0.0
Tokenizers 0.21.4