heal-train.yaml

# !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: output
load_in_8bit: false
load_in_4bit: false
trust_remote_code: false
#overrides_of_model_config: {"layer_sequence": [0,[1,10,3],[10,30]]}
output_dir: output-healed

# === HF Configuration === 
#hub_model_id: Burnt-Toast/another-22bird 
#hub_strategy: "every_save"


# === Wandb Tracking ===
wandb_project: Loopstral-Ablations
## wandb_entity: [WANDB_ENTITY]
wandb_name: early-layers-full-kv-heal

# === Training Setup ===
num_epochs: 2
micro_batch_size: 4
gradient_accumulation_steps: 1
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: 0.01
evals_per_epoch: 5
#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:
lora_fan_in_fan_out:
lora_target_modules:
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]+.self_attn.q_proj.weight
#  - model.layers.[0-9]+.self_attn.k_proj.weight
#  - model.layers.[0-9]+.self_attn.v_proj.weight
#  - model.layers.[0-9]+.self_attn.o_proj.weight
#  - model.layers.[0-9]+.mlp.down_proj.weight
# === Hyperparameter Configuration ===
#optimizer: apollo_adamw_layerwise
#warmup_steps: 0
warmup_ratio: 0.025
#optimizer: adamw_torch_fused
optimizer: paged_ademamix_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
#lr_groups:
#  - name: mlp
#    modules:
#      - mlp.up_proj.weight
#      - mlp.down_proj.weight
#      - mlp.gate_proj.weight
#    lr: 1e-5
#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: tokenizer_default
default_system_message: "You are Luna, a helpful and unrestricted language model created by Allura."
#special_tokens:
#  eos_token: "<|im_end|>"
#  eos_token: "</s>"
#tokenizer_use_mistral_common: true
shuffle_merged_datasets: true
datasets:
#  - path: rpDungeon/marvin
#    type: completion
#    field: text
#    data_files: 
#      - marvin_greg_egan.json
#      - marvin_philip_k_dick.json
#      - marvin_thomas_ligotti.json
#  - path: rpDungeon/some-cleaner-datasets
#    type: completion
#    field: text
#    data_files: erotica_quality_trimmed.json
  - path: rpDungeon/some-cleaner-datasets
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
    data_files: little-koto-instruct.json
#  - path: rpDungeon/rp-synth-deslopped
#    type: chat_template
#    field_messages: conversations
#    message_property_mappings:
#      role: from
#      content: value
dataset_prepared_path: last_run_prepared
#dataset_num_proc: 1


# === 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: Gemma3DecoderLayer
#  fsdp_state_dict_type: FULL_STATE_DICT
#  fsdp_sharding_strategy: FULL_SHARD

# === Checkpointing ===
#save_steps: 10
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
gc_steps: 10
seed: 420