metadata

license: mit
library_name: peft
tags:
  - generated_from_trainer
base_model: microsoft/phi-2
model-index:
  - name: phi-2-prompt-injection-QLoRA
    results: []
datasets:
  - HuggingFaceH4/no_robots
  - Dahoas/synthetic-hh-rlhf-prompts
  - HuggingFaceH4/ultrachat_200k
  - Lakera/gandalf_ignore_instructions
  - imoxto/prompt_injection_cleaned_dataset-v2
  - hackaprompt/hackaprompt-dataset
  - rubend18/ChatGPT-Jailbreak-Prompts
  - HuggingFaceH4/instruction-dataset

phi-2-prompt-injection-QLoRA

Weights updated at 03/07/2024. (Training epochs increased, accuracy improved than before)

View training code: https://github.com/AIM-Intelligence/phi-2-prompt-injection-QLoRA

Try out the model in: https://aim-intelligence.com

This model is a fine-tuned version of microsoft/phi-2 on the None dataset. It achieves the following results on the evaluation set:

eval_loss: 0.0000
eval_precision: 1.0
eval_recall: 1.0
eval_f1-score: 1.0
eval_accuracy: 1.0
eval_runtime: 16.0258
eval_samples_per_second: 8.424
eval_steps_per_second: 1.061
step: 0

Model description

More information needed

Intended uses & limitations

tokenizer = AutoTokenizer.from_pretrained("ysy970923/phi-2-prompt-injection-QLoRA")
model = AutoModelForSequenceClassification.from_pretrained("ysy970923/phi-2-prompt-injection-QLoRA", load_in_4bit=True, torch_dtype=torch.bfloat16, id2label={0: "SAFE", 1: "INJECTION"})
# LABEL_0 is safe, LABEL_1 is prompt_injection

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
train_batch_size: 2
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 8
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: constant
lr_scheduler_warmup_ratio: 0.03
num_epochs: 3.0

Framework versions

PEFT 0.8.2
Transformers 4.38.1
Pytorch 2.2.0+cu121
Datasets 2.17.1
Tokenizers 0.15.2