See axolotl config

axolotl version: 0.15.0.dev0

# === Model Configuration ===
base_model: rpDungeon/new-apertus-12b-untrained
load_in_8bit: false
load_in_4bit: false

# === HF Configuration === 
hub_model_id: BirdToast/new-apertus-12b-s1
hub_strategy: "every_save"
output_dir: ckpts

# === Wandb Tracking ===
wandb_project: ApertusV4
# wandb_entity: [WANDB_ENTITY]
wandb_name: 12b-cpt-s1

# === Training Setup ===
num_epochs: 1
micro_batch_size: 4
gradient_accumulation_steps: 1
sequence_len: 512
#sequence_parallel_degree: 2
#heads_k_stride: 1
sample_packing: true
#pad_to_sequence_len: true
#temperature: 0.7
#max_steps: 10
# === Evaluation ===
#val_set_size: 100
#evals_per_epoch: 10
#eval_steps: 20
#max_steps: 60
#eval_table_size:
#eval_max_new_tokens: 128
#eval_sample_packing: true
eval_strategy: "no"

# === LoRA Configuration ===
adapter:
lora_model_dir:
lora_r: 128
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_target_modules:
#  - up_proj
#  - down_proj
#  - gate_proj
#  - q_proj
#  - v_proj
#  - k_proj
#  - o_proj
#  - input_layernorm
#  - post_attention_layernorm
#  - embed_tokens
#  - lm_head

lora_fan_in_fan_out:
peft_use_rslora: true
lora_modules_to_save:
#  - embed_tokens
#  - lm_head
#fix_untrained_tokens: true
#lora_mlp_kernel: true
#lora_qkv_kernel: true
#lora_o_kernel: true
unfrozen_parameters:
  - model.layers.(2[4-9]|3[0-9]).*
#  - model.layers.[0-9+].mlp.up_proj
#  - model.layers.[0-9+].mlp.down_proj
#  - model.layers.[0-9+].feedforward_layernorm
#  - embed_tokens
#  - lm_head
# === Hyperparameter Configuration ===
#optimizer: apollo_adamw_layerwise
#warmup_steps: 0
warmup_ratio: 0.025
#optimizer: adamw_torch_fused
optimizer: came_pytorch
optim_args:
  enable_stochastic_rounding: true
  enable_cautious: true
  enable_cautious_weight_decay: true
#  enable_8bit: true
# Apollo-mini configuration:
#optim_args: "proj=random,rank=128,scale=128.0,scale_type=tensor,update_proj_gap=100"
# Regular Apollo configuration:
# optim_args: 
#optim_target_modules: all_linear
learning_rate: 1e-4
lr_scheduler: cosine
#cosine_min_lr_ratio: 0.2
#lr_scheduler: cosine_with_min_lr
#lr_scheduler_kwargs:
#  cosine_min_lr: 1e-6
weight_decay: 0.01
max_grad_norm: 1.0
#warmup_steps: 0
#warmup_ratio: 0.025


# === Data Configuration ===
#
#chat_template: jinja
chat_template: chatml
special_tokens:
#  eos_token: "<|im_end|>"
#  eos_token: "</s>"
#tokenizer_use_mistral_common: true
shuffle_merged_datasets: true
datasets:
  - path: rpDungeon/marvin
    data_files: marvin_full_text_only.json
    type: completion

dataset_prepared_path: last_run_prepared


# === Plugins ===
plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

# === Hardware Optimization ===
gradient_checkpointing: true
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
#liger_fused_linear_cross_entropy: true
cut_cross_entropy: true

#deepspeed: ../axolotl/deepspeed_configs/zero2.json

# === Checkpointing ===
#save_steps: 2
saves_per_epoch: 1
save_total_limit: 1

# === Advanced Settings ===
bf16: auto
flash_attention: true
train_on_inputs: false
group_by_length: false
save_safetensors: true
logging_steps: 1
seed: 69
gc_steps: 10

new-apertus-12b-s1

This model is a fine-tuned version of rpDungeon/new-apertus-12b-untrained on the rpDungeon/marvin 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: 0.0001
• train_batch_size: 4
• eval_batch_size: 4
• seed: 69
• 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: 68
• training_steps: 2730

Training results

Framework versions

• Transformers 4.57.3
• Pytorch 2.8.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2