base_model: mistralai/Mistral-7B-Instruct-v0.3
# 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: AiAF/QLoRA-Finetune-LMTV-TM-9-2320-365-10

load_in_8bit: false
load_in_4bit: true

datasets:
  - path: AiAF/TM_plain_qa_list_with_special_tokens.jsonl
    ds_type: json
    type: chat_template
    chat_template: chatml
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    roles:
      user:
        - human
      assistant:
        - gpt
      system:
        - system

dataset_prepared_path: last_run_prepared
val_set_size: 0.1
output_dir: ./TM-9-2320-365-10_V11

save_total_limit: 100

adapter: qlora
lora_model_dir:

sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

adapter: qlora
lora_r: 256
lora_alpha: 512
lora_dropout: 0.05
#lora_target_linear: true
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj

wandb_project: "LLM-Pretraining"
wandb_watch: "all"
wandb_name: "LMTV-TM-V11"
wandb_log_model: "false"
wandb_run_id: "LMTV-TM-V11"

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 30
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.000005

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
evals_per_epoch: 5
saves_per_epoch: 2
weight_decay: 0.0
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"