base_model: google/gemma-3n-E2B-it

# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
cut_cross_entropy: true

load_in_8bit: false
load_in_4bit: true

# for use with fft to only train on language model layers
# unfrozen_parameters:
  # - model.language_model.*
  # - lm_head
  # - embed_tokens


chat_template: gemma3n
eot_tokens:
  - <end_of_turn>
datasets:
  - path: cgato/SlimOrcaDedupCleaned
    type: chat_template
    split: train[:1%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value

val_set_size: 0.0
output_dir: ./outputs/out

adapter: qlora
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
# lora_target_linear: # Does not work with gemma3n currently
lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'

sequence_len: 2048
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
resume_from_checkpoint:
logging_steps: 1
# flash_attention: true  # Any attention impl does not work with gemma3n now

warmup_ratio: 0.1
evals_per_epoch:
saves_per_epoch: 1
weight_decay: 0.0
special_tokens: