Harryxun/llama-pretrained

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Nov 13, 2025License:llama3.2Architecture:Transformer Featherless Exclusive Warm

The Harryxun/llama-pretrained model is a 1 billion parameter language model, fine-tuned from Meta Llama-3.2-1B. This model was trained using specific hyperparameters including a learning rate of 0.0005 and cosine learning rate scheduler over 2 epochs. While its specific training dataset and intended uses are not detailed, it represents a base model for further fine-tuning or research into Llama-3.2 architecture at a smaller scale.

Loading preview...

Model Overview

The Harryxun/llama-pretrained model is a 1 billion parameter language model, derived from the meta-llama/Llama-3.2-1B architecture. This model has undergone a fine-tuning process, though the specific dataset used for this training is not disclosed in the available documentation.

Training Details

The fine-tuning procedure involved several key hyperparameters:

  • Learning Rate: 0.0005
  • Batch Sizes: train_batch_size of 32, eval_batch_size of 32, with gradient_accumulation_steps set to 8, resulting in a total_train_batch_size of 256.
  • Optimizer: adamw_torch_fused with default betas and epsilon.
  • Scheduler: Cosine learning rate scheduler with 1000 warmup steps.
  • Epochs: Trained for 2 epochs.
  • Precision: Utilized Native AMP for mixed-precision training.

Current Status and Limitations

As of the current documentation, further information regarding the model's specific intended uses, limitations, and the dataset it was trained on is needed. This model serves as a fine-tuned base, potentially suitable for researchers or developers looking to experiment with the Llama-3.2 architecture at a 1 billion parameter scale, with the understanding that its specific capabilities and performance characteristics are yet to be fully detailed.