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

Model	Max Global VRAM (MB)	Max Process VRAM (MB)	Max Reserved VRAM (MB)	Max Allocated VRAM (MB)
Llama-3.1-8B	18521.98	16630.42	16196.30	16060.54
Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0	16319.97	14428.41	13994.30	13879.42

Inference Mode Latency Metrics

Model	Latency Mean (s)	Throughput (tokens/s)
Llama-3.1-8B	0.8104	38.2536
Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0	0.5530	56.0570

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

Evaluation Data and Process:

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

Additional Benchmark Results

BoolQ 0-shots Benchmark Results

Model	Average Score	boolq (0 shots)	boolq contrastset (0 shots)
meta-llama/Meta-Llama-3.1-8B	0.569	0.569	0.568
Na0s/Llama-3.1-8B-Pruned-4-Layers	0.240	0.240	0.240
Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0	0.833	0.834	0.831

BigBench 0-shots Benchmark Results

Model	Average Score	bigbench:causal_judgment (0 shots)	bigbench:date_understanding (0 shots)	bigbench:disambiguation_qa (0 shots)	bigbench:geometric_shapes (0 shots)	bigbench:logical_deduction (0 shots)	...
meta-llama/Meta-Llama-3.1-8B	0.351	0.574	0.499	0.302	0.164	0.208	...
Na0s/Llama-3.1-8B-Pruned-4-Layers	0.299	0.537	0.341	0.314	0.200	0.212	...
Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0	0.364	0.579	0.610	0.407	0.264	0.208	...

Few Shots Benchmark Results

Model	Average Score	arc:challenge (25 shots)	hellaswag (10 shots)	mmlu:abstract_algebra (5 shots)	mmlu:college_chemistry (5 shots)	mmlu:college_computer_science (5 shots)	mmlu:college_mathematics (5 shots)	...
meta-llama/Meta-Llama-3.1-8B	0.552	0.541	0.620	0.290	0.450	0.480	0.350	...
Na0s/Llama-3.1-8B-Pruned-4-Layers	0.516	0.462	0.549	0.290	0.440	0.460	0.280	...
Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0	0.544	0.479	0.554	0.340	0.480	0.520	0.350	...

BigBench 3-shots Benchmark Results

Model	Average Score	bigbench:causal_judgment (3 shots)	bigbench:date_understanding (3 shots)	bigbench:disambiguation_qa (3 shots)	bigbench:geometric_shapes (3 shots)	bigbench:logical_deduction (3 shots)	...
meta-llama/Meta-Llama-3.1-8B	0.442	0.563	0.596	0.593	0.181	0.298	...
Na0s/Llama-3.1-8B-Pruned-4-Layers	0.420	0.563	0.642	0.574	0.217	0.258	...
Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0	0.450	0.621	0.686	0.663	0.225	0.332	...

Overall Average Score

Model	Overall Average Score
meta-llama/Meta-Llama-3.1-8B	0.472
Na0s/Llama-3.1-8B-Pruned-4-Layers	0.364
Na0s/Llama-3.1-8B-Pruned-4-Layers_LoRA-PEFT-3.0	0.513

Environmental Impact:

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