Weni/WeniGPT-Mistral-7B-instructBase-4bit

Training procedure

The following bitsandbytes quantization config was used during training:

• quant_method: bitsandbytes
• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: bfloat16

Framework versions

• PEFT 0.4.0

This model was trained with this parameters:

max_seq_length = 2048

training_arguments_mistral = {
    'num_train_epochs':10,
    'per_device_train_batch_size':2,
    'gradient_accumulation_steps':2,
    'gradient_checkpointing':True,
    'optim':'adamw_torch',
    'lr_scheduler_type':'constant_with_warmup',
    'logging_steps':10,
    'evaluation_strategy':'epoch',
    'save_strategy':"epoch",
    'load_best_model_at_end':True,
    'learning_rate':4e-4,
    'save_total_limit':3,
    'fp16':True,
    'tf32': True, 
    'max_steps':8000,
    'max_grad_norm':0.3,
    'warmup_ratio':0.03,
    'disable_tqdm':False,
    'weight_decay':0.001,
    'hub_model_id':'Weni/WeniGPT-Mistral-7B-instructBase-4bit',
    'push_to_hub':True,
    'hub_strategy':'every_save',
    'hub_token':token,
    'hub_private_repo':True,