cygu/llama-2-7b-sampling-watermark-distill-kth-shift256

Model description

Sampling-based watermark distilled Llama 2 7B using the KTH \\(s=256\\) watermarking strategy in the paper On the Learnability of Watermarks for Language Models.

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 1e-05
• —trainbatchsize: 8
• —evalbatchsize: 8
• —seed: 42
• —distributed_type: multi-GPU
• —num_devices: 4
• —gradientaccumulationsteps: 4
• —totaltrainbatch_size: 128
• —totalevalbatch_size: 32
• —optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• —lrschedulertype: cosine
• —lrschedulerwarmup_steps: 500
• —num_epochs: 1.0

Framework versions

• —Transformers 4.29.2
• —Pytorch 2.0.1+cu117
• —Datasets 2.13.1
• —Tokenizers 0.13.3