AlexHung29629/my_sft_checkpoints

See axolotl config

axolotl version: 0.10.0.dev0

base_model: /mnt/shared/tp1-an1/alex/Magistral/orignal_model

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

datasets:
  - path: AlexHung29629/merged_aime_yamol_dataset
    split: train
    type:
      system_prompt: ""
      field_system: system
      field_instruction: input
      field_output: output
      format: "{instruction}"
      no_input_format: "{instruction}"
      
dataset_prepared_path: ./sft_dataprep/
val_set_size: 0
output_dir: ./placeholder_sft_4ep/
shuffle_merged_datasets: false

sequence_len: 32768
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

wandb_project: TP1_2025_05
wandb_entity:
wandb_watch:
wandb_name: Mistral-24B-SFT-250624
use_tensorboard: true

save_only_model: true
gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 5e-6
max_grad_norm: 1.0

adam_beta1: 0.9
adam_beta2: 0.95
adam_epsilon: 1e-8

bf16: true
tf32: false

logging_steps: 1
flash_attention: true
xformers_attention: false
sdp_attention: false

warmup_ratio: 0.05
saves_per_epoch: 1
weight_decay: 0

fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: true
  fsdp_cpu_ram_efficient_loading: true
  fsdp_activation_checkpointing: true
  fsdp_transformer_layer_cls_to_wrap: MistralDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP

seed: 42
auto_resume_from_checkpoints: true

placeholder_sft_4ep/

This model was trained from scratch on the AlexHung29629/merged_aime_yamol_dataset 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: 5e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 16
• total_train_batch_size: 16
• total_eval_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 4
• num_epochs: 4.0

Training results

Framework versions

• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.5.1
• Tokenizers 0.21.1