base_model: allura-org/Gemma-3-Glitter-12B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets: 
  - path: NewEden/LIMARP-Complexity
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/Creative_Writing-Complexity
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/wata-Oshi-No-prefix
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/No-Prefix-LN
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/Storium-No-Prefix
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/PIPPA-Mega-Filtered
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
  - path: NewEden/OpenCAI-ShareGPT
    type: chat_template
    roles_to_train: ["gpt"]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    train_on_eos: turn
val_set_size: 0.01
output_dir: ./outputs

sequence_len: 16384
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true


plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: false
cut_cross_entropy: true
optimizer: apollo_adamw
optim_args: proj=random,rank=1,scale=128.0,scale_type=tensor,update_proj_gap=200
optim_target_modules:
  - .*.attn.*
  - .*.mlp.*
gradient_checkpointing: unsloth
flash_attention: true

adapter: lora
lora_model_dir:
lora_r: 64
lora_alpha: 32
lora_dropout: 0.1
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 4
lr_scheduler: rex
learning_rate: 1e-5
weight_decay: 0.02

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

early_stopping_patience:
resume_from_checkpoint:
#auto_resume_from_checkpoints: true
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 35
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1

debug:
deepspeed: 
fsdp:
fsdp_config: