#Note that we are switching from the regular chat template to chatml.
#If you experience problems with the special tokens, training for more epochs can help.
#After training, merge the model before inference otherwise you might
#face problems with the special tokens.

base_model: mistralai/Mistral-7B-Instruct-v0.2
# optionally might have model_type or tokenizer_type
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

load_in_8bit: false
load_in_4bit: true

chat_template: chatml
rl: dpo
datasets:
  - path: olivermolenschot/alpaca_messages_dpo_test
    type: chat_template.default
    field_messages: conversation
    field_chosen: chosen
    field_rejected: rejected
    message_property_mappings:
      role: role
      content: content

dataset_prepared_path:
val_set_size: 0.05
output_dir: ./outputs/dpo-qlora

sequence_len: 2048
sample_packing: false
pad_to_sequence_len: true

adapter: qlora
lora_model_dir:
lora_r: 8
lora_alpha: 16
lora_dropout: 0.2
lora_target_linear: true

lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj
lora_modules_to_save:
 - embed_tokens
 - lm_head

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 16
num_epochs: 6
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0001

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: false

warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.0
special_tokens:
  bos_token: "<|im_start|>"
  eos_token: "<|im_end|>"