UsernameJustAnother/Nemo-12B-Marlin-v5

Uploaded model

• Developed by: UsernameJustAnother
• License: apache-2.0
• Finetuned from model : unsloth/Mistral-Nemo-Instruct-2407

I am a terrible liar. I came across another dataset I had to use, and this is the result. Still experimental, as I made these to teach myself the basics of fine-tuning, with notes extensively borrowed from https://huggingface.co./nothingiisreal/MN-12B-Celeste-V1.9

It is an RP finetune using 10,801 human-generated conversations of varying lengths from a variety of sources and curated by me, trained in ChatML format.

The big differences from Celeste is a different LoRA scaling factor. Celeste uses 8; I did several tests with this data before concluding I got lower training loss with 2.

Training took around 5 hours on a single Colab A100 (but I didn't do an eval loop). Neat that I could get it all to fit into 40GB of vRAM thanks to Unsloth.

It was trained with the following settings:

==((====))==  Unsloth - 2x faster free finetuning | Num GPUs = 1
   \\   /|    Num examples = 10,801 | Num Epochs = 2
O^O/ \_/ \    Batch size per device = 2 | Gradient Accumulation steps = 4
\        /    Total batch size = 8 | Total steps = 2,700
 "-____-"     Number of trainable parameters = 912,261,120

[ 14/2700 01:20 < 4:59:21, 0.15 it/s, Epoch 0.01/2] 
[2040/2040 3:35:30, Epoch 2/2] 

model = FastLanguageModel.get_peft_model(
    model,
    r = 256,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 32,  #   32 / sqrt(256) gives a scaling factor of 2
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = True,  # setting the adapter scaling factor to lora_alpha/math.sqrt(r) instead of lora_alpha/r
    loftq_config = None, # And LoftQ
)

lr_scheduler_kwargs = {
    'min_lr': 0.0000024  # Adjust this value as needed
}

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = train_ds,
    compute_metrics = compute_metrics,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 2,
    packing = False, # Can make training 5x faster for short sequences.
    args = TrainingArguments(
        per_device_train_batch_size = 2,
        per_device_eval_batch_size = 2, # defaults to 8!
        gradient_accumulation_steps = 4,
        warmup_steps = 5,
        num_train_epochs = 2,
        learning_rate = 8e-5,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        fp16_full_eval = True, # stops eval from trying to use fp32
        eval_strategy = "no", # 'no', 'steps', 'epoch'. Don't use this without an eval dataset etc
        eval_steps = 1, # is eval_strat is set to 'steps', do every N steps.
        logging_steps = 1, # so eval and logging happen on the same schedule
        optim = "adamw_8bit", 
        weight_decay = 0.01,
        lr_scheduler_type = "cosine_with_min_lr", # linear, cosine, cosine_with_min_lr, default linear
        lr_scheduler_kwargs = lr_scheduler_kwargs, # needed for cosine_with_min_lr
        seed = 3407,
        output_dir = "outputs",
    ),
)

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