nbeerbower/mistral-nemo-narwhal-12B

Mahou-1.5-mistral-nemo-12B-lorablated finetuned on reddit-dpo.

Method

ORPO tuned with 8x A100 for 1 epoch.

QLoRA config:

# QLoRA config
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch_dtype,
    bnb_4bit_use_double_quant=True,
)
# LoRA config
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=['up_proj', 'down_proj', 'gate_proj', 'k_proj', 'q_proj', 'v_proj', 'o_proj']
)

Training config:

orpo_args = ORPOConfig(
    run_name=new_model,
    learning_rate=8e-6,
    lr_scheduler_type="linear",
    max_length=2048,
    max_prompt_length=1024,
    max_completion_length=1024,
    beta=0.1,
    per_device_train_batch_size=4,
    per_device_eval_batch_size=4,
    gradient_accumulation_steps=1,
    optim="paged_adamw_8bit",
    num_train_epochs=2,
    evaluation_strategy="steps",
    eval_steps=0.2,
    logging_steps=1,
    warmup_steps=10,
    max_grad_norm=10,
    report_to="wandb",
    output_dir="./results/",
    bf16=True,
    gradient_checkpointing=True,
)