smohammadi/qated-nvfp4-llama3B

See axolotl config

axolotl version: 0.13.0.dev0

base_model: meta-llama/Llama-3.2-3B
hub_model_id: smohammadi/qat-nvfp4-llama3B

load_in_8bit: false
load_in_4bit: false
strict: false

  #chunked_cross_entropy: true
  #plugins:
  #  - axolotl.integrations.liger.LigerPlugin

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_out/
dataset_prepared_path: ./outputs/qat_out/dataset_prepared

sample_packing: false #true
sequence_len: 4096
flash_attention: true
  #flex_attention: true
  #flex_attn_compile_kwargs:
  # dynamic: false
  #mode: max-autotune-no-cudagraphs
quantization:
  activation_dtype: nvfp4
  weight_dtype: nvfp4
  group_size: 16 # only group_size of 16 is supported with nvfp4

wandb_project: qat_v2
wandb_entity: 
wandb_watch:
wandb_name: bf16-nvfp4
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 64
num_epochs: 1
optimizer: adamw_torch_fused
gradient_checkpointing: true

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

resume_from_checkpoint:
logging_steps: 1

evals_per_epoch: 1
saves_per_epoch: 1

warmup_ratio: 0.1
weight_decay: 0.0

special_tokens:
  pad_token: <|finetune_right_pad_id|>

# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

qat-nvfp4-llama3B

This model is a fine-tuned version of meta-llama/Llama-3.2-3B 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: 64
• eval_batch_size: 64
• seed: 42
• 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: cosine
• lr_scheduler_warmup_steps: 76
• training_steps: 769

Training results

Framework versions

• Transformers 4.55.4
• Pytorch 2.8.0+cu128
• Datasets 4.0.0
• Tokenizers 0.21.4