base_model: mistralai/Voxtral-Mini-3B-2507
processor_type: VoxtralProcessor

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

# Enable to use mistral-common tokenizer
tokenizer_use_mistral_common: true

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

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

load_in_4bit: true

# these 3 lines are needed for now to handle vision chat templates w images
skip_prepare_dataset: true
remove_unused_columns: false
sample_packing: false

# gemma3 doesn't seem to play nice with ddp
ddp_find_unused_parameters: true

eot_tokens:
  - <end_of_turn>

# sample dataset below requires downloading audio/image in advance
# wget https://huggingface.co/datasets/Nanobit/text-audio-2k-test/resolve/main/En-us-African_elephant.oga
datasets:
  - path: NanoBit/text-audio-2k-test
    type: chat_template
dataset_prepared_path:
val_set_size: 0.01
output_dir: ./outputs/out

adapter: qlora
lora_model_dir:

sequence_len: 2048
pad_to_sequence_len: false

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: true
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
logging_steps: 1
flash_attention: true

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