# !pip install transformers==4.55.4
# !pip install --no-deps trl==0.22.2
# !pip install --no-build-isolation mamba_ssm==2.2.5
# !pip install --no-build-isolation causal_conv1d==1.5.2
# === Model Configuration ===
base_model: LatitudeGames/Muse-12B
load_in_8bit: false
load_in_4bit: true

# === HF Configuration === 
hub_model_id: ToastyPigeon/muse-marvin-attn-lora
hub_strategy: "every_save"
output_dir: ckpts-mmarv

# === Training Setup ===
num_epochs: 1
micro_batch_size: 1
gradient_accumulation_steps: 4
sequence_len: 16384
#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: 0.025
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: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.1
lora_target_linear:
lora_target_modules:
  - q_proj
  - v_proj
  - k_proj
  - o_proj
lora_fan_in_fan_out:
peft_use_rslora: false
#lora_modules_to_save:
#  - embed_tokens
#  - lm_head
#fix_untrained_tokens: true
#lora_mlp_kernel: true
#lora_qkv_kernel: true
#lora_o_kernel: true

# === 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: 1e-5
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: 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: ToastyPigeon/gutenberg-sft
#    type: customchatml-regex
#    type: chat_template
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
#  - path: ToastyPigeon/SpringDragon
#    type: customcompletion-regex
#    type: completion
#    split: train
#  - path: ToastyPigeon/some-erotica
#    type: customcompletion-regex
#    type: completion
#    split: train[:10%]

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/zero3_bf16_cpuoffload_params.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: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: MistralDecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_sharding_strategy: FULL_SHARD
#  fsdp_version: 2
# === Wandb Tracking ===
wandb_project: MuseMarvin
# wandb_entity: [WANDB_ENTITY]
# wandb_name: [WANDB_RUN_NAME]

# === Checkpointing ===
#save_steps: 10
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
gc_steps: 10
seed: 69

muse-marvin-attn-lora

This model is a fine-tuned version of LatitudeGames/Muse-12B on the grimulkan/LimaRP-augmented, the ToastyPigeon/steve-and-marvin and the ToastyPigeon/kimi-stories-completion datasets. It achieves the following results on the evaluation set:

Loss: 2.4268
Memory/max Active (gib): 5.02
Memory/max Allocated (gib): 4.89
Memory/device Reserved (gib): 6.64

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: 1e-05
train_batch_size: 1
eval_batch_size: 1
seed: 69
distributed_type: multi-GPU
num_devices: 2
gradient_accumulation_steps: 4
total_train_batch_size: 8
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: 5
training_steps: 232

Training results

Training Loss	Epoch	Step	Validation Loss	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	2.5323	8.04	6.73	8.36
2.5888	0.1032	24	2.4883	5.02	4.89	6.64
2.4142	0.2065	48	2.4537	5.02	4.89	6.64
2.3697	0.3097	72	2.4418	5.02	4.89	6.64
2.2986	0.4129	96	2.4354	5.02	4.89	6.64
2.5054	0.5161	120	2.4314	5.02	4.89	6.64
2.6863	0.6194	144	2.4290	5.02	4.89	6.64
2.3196	0.7226	168	2.4277	5.02	4.89	6.64
2.3422	0.8258	192	2.4271	5.02	4.89	6.64
2.5976	0.9290	216	2.4268	5.02	4.89	6.64

Framework versions

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