# === Model Configuration ===
#base_model: apertus-12b-trained/cpt-part1-allmods/merged
base_model: allura-forge/apertus-12b-cpt-attempt
load_in_8bit: false
load_in_4bit: false

# === HF Configuration === 
hub_model_id: allura-forge/apertus-12b-cpt-s2-lora-2
hub_strategy: "every_save"
output_dir: apertus-12b-trained/cpt-part2-mlp-2

# === Wandb Tracking ===
wandb_project: ApertusV2
# wandb_entity: [WANDB_ENTITY]
wandb_name: 12b-cpt-part2

# === Training Setup ===
num_epochs: 1
micro_batch_size: 1
gradient_accumulation_steps: 16
sequence_len: 8192
#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: 200
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
lora_model_dir:
lora_r: 64
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear:
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.[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: paged_adamw_8bit
#optim_args:
#  enable_stochastic_rounding: true
#  enable_cautious: 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: 5e-6
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: allura-org/the-anarchist-library
    type: completion
    split: train[50%:]
#  - path: grimulkan/LimaRP-augmented
#    type: chat_template
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#  - path: allenai/tulu-3-sft-personas-instruction-following
#    type: chat_template
#    split: train[:10%]
#  - path: ToastyPigeon/mixed-medical-reasoning-formatted
#    type: chat_template
#    data_files: mixed-medical-thinking.json
#    split: train[:10%]
#  - path: ToastyPigeon/steve-and-marvin
#    type: completion
#    data_files: marvin.json
#  - path: ToastyPigeon/kimi-stories-completion
#    type: completion
#  - path: ToastyPigeon/new-story-dataset
 #   type: customcompletion-regex
#    type: completion
#    data_files: new-story-dataset-v2.json
#  - path: allura-org/fujin-instruct-v2
#    type: customchatml-regex
#    type: chat_template
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#  - path: ToastyPigeon/some-rp-extended
 #   type: customchatml-regex
#    type: chat_template
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#    roles_to_train: ["user","assistant"]
#  - path: Alfitaria/rosier-inf
#    type: completion
#    split: train[70%:]
#  - path: allura-forge/koto-instruct-sft-nothink
#    type: customchatml-regex
#    type: chat_template
#    split: train[:50%]
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#  - path: ToastyPigeon/SpringDragon
#    type: customcompletion-regex
#    type: completion
#    split: train
#  - path: ToastyPigeon/erotic-books-clone
#    type: customcompletion-regex
#    type: completion
#    split: train[:50%]
#    split: train[35%:45%]
#  - path: ToastyPigeon/tulu-mini
#    type: chat_template
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

# === FSDP Config === 
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_activation_checkpointing: true
  fsdp_use_orig_params: true
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: ApertusDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD
#fsdp_stage: 2
#fsdp_final_state_dict_type: FULL_STATE_DICT

# === Checkpointing ===
#save_steps: 2
saves_per_epoch: 10
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: 420
gc_steps: 10

apertus-12b-cpt-s2-lora-2

This model is a fine-tuned version of allura-forge/apertus-12b-cpt-attempt on the allura-org/the-anarchist-library dataset. It achieves the following results on the evaluation set:

Loss: 2.2376
Memory/max Active (gib): 6.89
Memory/max Allocated (gib): 6.88
Memory/device Reserved (gib): 8.47

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: 420
distributed_type: multi-GPU
num_devices: 2
gradient_accumulation_steps: 16
total_train_batch_size: 32
total_eval_batch_size: 2
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: 6
training_steps: 240

Training results

Training Loss	Epoch	Step	Validation Loss	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	2.2445	6.88	6.87	8.43
2.3237	0.0999	24	2.2437	6.89	6.88	8.47
2.0934	0.1999	48	2.2418	6.89	6.88	8.47
2.1321	0.2998	72	2.2407	6.89	6.88	8.88
2.2765	0.3998	96	2.2397	6.89	6.88	8.88
2.2691	0.4997	120	2.2387	6.89	6.88	8.47
2.1086	0.5997	144	2.2384	6.89	6.88	8.47
2.126	0.6996	168	2.2379	6.89	6.88	8.88
2.2464	0.7996	192	2.2375	6.89	6.88	8.47
2.2584	0.8995	216	2.2375	6.89	6.88	8.88
2.2255	0.9995	240	2.2376	6.89	6.88	8.47

Framework versions

PEFT 0.17.1
Transformers 4.56.1
Pytorch 2.7.1+cu126
Datasets 4.0.0
Tokenizers 0.22.1