Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0

<a href="https://ibb.co/xGnXR5t"><img src="https://i.ibb.co/P42FfCv/DALL-E-2024-08-08-05-21-39-An-artistic-representation-for-a-model-card-featuring-an-abstract-and-sty.webp" alt="DALL-E-2024-08-08-05-21-39-An-artistic-representation-for-a-model-card-featuring-an-abstract-and-sty" border="0"></a>

Model Card for Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0:

Model Details:

Model Description:

• Finetuned from model: Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-2.0 on teknium/openhermes.
• We pruned the 4 layers of meta-llama/Meta-Llama-3.1-8B that had the less impact on the performance of the model according to the paper The Unreasonable Ineffectiveness of the Deeper Layers.
• We have therefore 1.09B parameters less than the foundation model, which means less memory needed, faster training and less latency during inference mode.
• We then recovered the performance loss induced by the pruning process by fine-tuning (from 0.2642 MMLU-Pro 0-shot to 0.3120), this step is called healing the pruned model.

Upcoming Work:

• More healing through SFT/DPO/TPO to see if we can get closer to the meta-llama/Meta-Llama-3.1-8B performance (which has an MMLU-Pro 0-shot of 0.3659 vs 0.3120 for our model). (In Progress)
• Compare the same exact process when applied to meta-llama/LLama-3.1-70B.

Training Details:

model = FastLanguageModel.get_peft_model(
model,
r = 4, 
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                  "gate_proj", "up_proj", "down_proj",],
lora_alpha = 4,
lora_dropout = 0.05, 
bias = "none",    

use_gradient_checkpointing = "unsloth", 
random_state = 3407,
use_rslora = False,  
loftq_config = None, 
)

from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported

trainer = SFTTrainer(
model = model,
tokenizer = tokenizer,
train_dataset = dataset,
dataset_text_field = "completion",
max_seq_length = max_seq_length,
dataset_num_proc = 2,
packing = False, 
args = TrainingArguments(
    per_device_train_batch_size = 10,
    gradient_accumulation_steps = 4,
    warmup_steps = 5,
    max_steps=5000,
    learning_rate = 2e-4,
    fp16 = not is_bfloat16_supported(),
    bf16 = is_bfloat16_supported(),
    logging_steps = 1,
    optim = "adamw_8bit",
    weight_decay = 0.01,
    lr_scheduler_type = "cosine",
    seed = 3407,
    output_dir = "outputs_4",
    push_to_hub=True,
    hub_always_push=True,
),
)

Training Data:

teknium/openhermes

Memory and Latency gain (Using Optimum-Benchmark):

Load Mode Memory Metrics

Inference Mode Latency Metrics

Evaluation:

• (Foundation model) MMLU Pro 0-shot of meta-llama/Meta-Llama-3.1-8B: 0.3659
• (Pruned model) MMLU Pro 0-shot of Na0s/Llama-3.1-8B-Pruned-4-Layers: 0.2642
• (Healed model) MMLU Pro 0-shot of Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0: 0.3120

<a href="https://ibb.co/6ngkh8w"><img src="https://i.ibb.co/dWGR9tM/Screenshot-2024-08-08-at-7-41-26-AM.png" alt="Screenshot-2024-08-08-at-7-41-26-AM" border="0" width=600 height=600></a>

Evaluation Data and Process:

• TIGER-AI-Lab/MMLU-Pro.
• HuggingFace Lighteval benchmarking repo.

Additional Benchmark Results

BoolQ 0-shots Benchmark Results

BigBench 0-shots Benchmark Results

Few Shots Benchmark Results

BigBench 3-shots Benchmark Results

Overall Average Score

Environmental Impact:

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).