AlexHung29629/gpt-oss-mb-v2-mxfp4

See axolotl config

axolotl version: 0.13.0.dev0

base_model: ./gpt-oss-20b
use_kernels: false
model_quantization_config: Mxfp4Config
model_quantization_config_kwargs:
  dequantize: true

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding
unfrozen_parameters:
  - \S+self_attn\S+

datasets:
  - path: AlexHung29629/mbpii
    type: chat_template

dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/gpt-oss-v2

dataloader_num_workers: 0
dataloader_pin_memory: True

sequence_len: 16384
sample_packing: true
eval_sample_packing: false
remove_unused_columns: false
pad_to_sequence_len: true

wandb_project: mb_pii
wandb_name: v2
tensorboard: true

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1

optimizer: adamw_torch_fused
lr_scheduler: constant_with_warmup
learning_rate: 2e-6
save_only_model: true
tensorboard: true
bf16: true
tf32: true

#flash_attention: true
eager_attention: true
#sdp_attention: true
#attn_implementation: kernels-community/vllm-flash-attn3
#flex_attention: true
#flex_attn_compile_kwargs:
#  dynamic: false
#  mode: max-autotune-no-cudagraphs

torch_compile: true

#gradient_checkpointing: true
#activation_offloading: true

logging_steps: 1
saves_per_epoch: 1

warmup_ratio: 0.05

special_tokens:
eot_tokens:
  - "<|end|>"
  - "<|return|>"

fsdp_version: 2
fsdp_config:
  offload_params: false
  state_dict_type: FULL_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: GptOssDecoderLayer
  reshard_after_forward: true
  activation_checkpointing: true

outputs/gpt-oss-v2

This model was trained from scratch on the AlexHung29629/mbpii 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-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 8
• total_eval_batch_size: 8
• 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_with_warmup
• training_steps: 17

Training results

Framework versions

• Transformers 4.55.0
• Pytorch 2.7.1+cu128
• Datasets 4.0.0
• Tokenizers 0.21.4