See axolotl config

axolotl version: 0.13.0.dev0

# !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-32k-lora
hub_strategy: "every_save"
output_dir: ckpts-mmarv

# === Wandb Tracking ===
wandb_project: MuseMarvin
# wandb_entity: [WANDB_ENTITY]
wandb_name: r32-qlora-32k

# === Training Setup ===
num_epochs: 2
micro_batch_size: 1
gradient_accumulation_steps: 4
sequence_len: 32768
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:
#  - up_proj
#  - down_proj
#  - gate_proj
  - q_proj
  - v_proj
  - k_proj
#  - o_proj
#  - input_layernorm
#  - post_attention_layernorm
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

# === Checkpointing ===
#save_steps: 10
saves_per_epoch: 2
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-32k-lora

This model is a fine-tuned version of LatitudeGames/Muse-12B on the ToastyPigeon/steve-and-marvin dataset. It achieves the following results on the evaluation set:

• Loss: 2.5071
• Memory/max Active (gib): 4.98
• Memory/max Allocated (gib): 4.89
• Memory/device Reserved (gib): 6.9

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: 10
• training_steps: 420

Training results

Training Loss	Epoch	Step	Validation Loss	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	2.6286	8.04	6.73	8.36
2.4233	0.0993	21	2.6047	4.98	4.89	6.9
2.5581	0.1986	42	2.5627	4.98	4.89	6.9
2.3368	0.2979	63	2.5447	4.98	4.89	6.9
2.5579	0.3972	84	2.5328	4.98	4.89	6.9
2.4241	0.4965	105	2.5253	4.98	4.89	6.9
2.4608	0.5957	126	2.5199	4.98	4.89	6.9
2.8143	0.6950	147	2.5156	4.98	4.89	6.9
2.6305	0.7943	168	2.5129	4.98	4.89	6.9
2.3989	0.8936	189	2.5105	4.98	4.89	6.9
2.6816	0.9929	210	2.5096	4.98	4.89	6.9
2.629	1.0898	231	2.5092	4.98	4.89	6.9
2.4645	1.1891	252	2.5088	4.98	4.89	6.9
2.3738	1.2884	273	2.5081	4.98	4.89	6.9
2.3651	1.3877	294	2.5076	4.98	4.89	6.9
2.4476	1.4870	315	2.5073	4.98	4.89	6.9
2.4091	1.5863	336	2.5072	4.98	4.89	6.9
2.6352	1.6856	357	2.5071	4.98	4.89	6.9
2.5311	1.7849	378	2.5071	4.98	4.89	6.9
2.5747	1.8842	399	2.5071	4.98	4.89	6.9
2.3871	1.9835	420	2.5071	4.98	4.89	6.9

Framework versions

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