TinyLlama-1.1B-Hinglish-LORA-v1.0 model

• Developed by: Kiran Kunapuli
• License: apache-2.0
• Finetuned from model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
• Model config:

    model = FastLanguageModel.get_peft_model(
    model,
    r = 64, 
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 128,
    lora_dropout = 0, 
    bias = "none",   
    use_gradient_checkpointing = True, 
    random_state = 42,
    use_rslora = True,  
    loftq_config = None, 
    )
  

• Training parameters:

    trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 2,
    packing = True, 
    args = TrainingArguments(
        per_device_train_batch_size = 12,
        gradient_accumulation_steps = 16,
        warmup_ratio = 0.1,
        num_train_epochs = 1,
        learning_rate = 2e-4,
        fp16 = not torch.cuda.is_bf16_supported(),
        bf16 = torch.cuda.is_bf16_supported(),
        logging_steps = 1,
        optim = "paged_adamw_32bit",
        weight_decay = 0.001,
        lr_scheduler_type = "cosine",
        seed = 42,
        output_dir = "outputs",
        report_to = "wandb",
      ),
    )
  

• Training details:

  ==((====))==  Unsloth - 2x faster free finetuning | Num GPUs = 1
     \\   /|    Num examples = 15,464 | Num Epochs = 1
  O^O/ \_/ \    Batch size per device = 12 | Gradient Accumulation steps = 16
  \        /    Total batch size = 192 | Total steps = 80
   "-____-"     Number of trainable parameters = 50,462,720
  
  GPU = NVIDIA GeForce RTX 3090. Max memory = 24.0 GB.
  Total time taken for 1 epoch - 2h:35m:28s
  9443.5288 seconds used for training.
  157.39 minutes used for training.
  Peak reserved memory = 17.641 GB.
  Peak reserved memory for training = 15.344 GB.
  Peak reserved memory % of max memory = 73.504 %.
  Peak reserved memory for training % of max memory = 63.933 %.
  

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

[NOTE] TinyLlama's internal maximum sequence length is 2048. We use RoPE Scaling to extend it to 4096 with Unsloth!