Model Card for Model ID
This model checkpoint is the TinyLlama-1.1B fine-tuned on alpaca dataset.
Model Details
Model Sources
- Repository: https://github.com/jzhang38/TinyLlama
- Paper: [https://arxiv.org/abs/2404.02406]
Uses
The use of this model should comply with the restrictions from TinyLlama-1.1b and Stanford Alpaca.
How to Get Started with the Model
Use the code below to get started with the model.
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("luckychao/TinyAlpaca-1.1B")
model = AutoModelForCausalLM.from_pretrained("luckychao/TinyAlpaca-1.1B")
Training Details
Training Data
We use the alpaca dataset, which is created by researchers from Stanford University.
Training Procedure
We follow the same training procedure and mostly same hyper-parameters to fine-tune the original Alpaca model on Llama. The procedure can be found in stanford_alpaca project.
Training Hyperparameters
--num_train_epochs 3 \
--per_device_train_batch_size 2 \
--per_device_eval_batch_size 2 \
--gradient_accumulation_steps 4 \
--evaluation_strategy "no" \
--save_strategy "steps" \
--save_steps 1000 \
--save_total_limit 1 \
--learning_rate 2e-5 \
--weight_decay 0. \
--warmup_ratio 0.03 \
--lr_scheduler_type "cosine" \
--logging_steps 1 \
--bf16 True \
--fsdp "full_shard auto_wrap" \
--fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
--model_max_length 2048
Citation
The model is mostly developed for the paper below. Please cite it if you find the repository helpful.
BibTeX:
@article{hao2024exploring,
title={Exploring Backdoor Vulnerabilities of Chat Models},
author={Hao, Yunzhuo and Yang, Wenkai and Lin, Yankai},
journal={arXiv preprint arXiv:2404.02406},
year={2024}
}