base_model: mistralai/Mistral-Small-3.2-24B-Instruct-2506
processor_type: AutoProcessor

load_in_8bit: false
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

chat_template: mistral_v7_tekken
datasets:
  - path: hardlyworking/HardlyRPv2-10k
    type: chat_template
    split: train
    field_messages: messages
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/out

adapter: qlora
lora_model_dir:

sequence_len: 8192
pad_to_sequence_len: false

lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 16
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: true
fp16:
tf32: true

gradient_checkpointing: unsloth
logging_steps: 1
# flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
sdp_attention: true

warmup_ratio: 0.1
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.0
special_tokens:

# save_first_step: true  # uncomment this to validate checkpoint saving works with your config