Llama-3-8B-spectrum-25
This model is a fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct on the yuvraj17/finetune_alpaca_1K dataset. It achieves the following results on the evaluation set:
- Loss: 1.2791
Spectrum Fine-tuning:
I have used the Spectrum Fine-tuning method as described in Eric Hartford et. al 2024, which selectively targets some t% of the model layers with the highest Signal-to-Noise Ratio (SNR). By focusing on the most information-dense layers, this approach maximizes fine-tuning efficiency while minimizing compute resources.
The key goal of Spectrum Fine-tuning is: minimize the memory footprint and accelerate LLM training without sacrificing performance.
The 25% layer selection ensures minimal computational overhead for fine-tuning.
Training:
- Trained on 2x A40s (48GB VRAM each) for over 1 hour using the Axolotl.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 4
- eval_batch_size: 4
- seed: 42
- distributed_type: multi-GPU
- num_devices: 2
- gradient_accumulation_steps: 4
- total_train_batch_size: 32
- total_eval_batch_size: 8
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 100
- num_epochs: 2
Framework versions
- Axolotl 0.4.1
- Transformers 4.44.2
- Pytorch 2.4.0+cu121
- Datasets 2.20.0
- Tokenizers 0.19.1