Home

Awesome

<div style="text-align:center"> <img src="https://www.cdeng.net/k2/k2.png" alt="k2-logo" width="200"/> <h2>🏔️ K2 (GeoLLaMA) Large Language Model for Geoscience</h2> </div>

<a href='https://arxiv.org/abs/2306.05064'><img src='https://img.shields.io/badge/Paper-ArXiv-C71585'></a> <a href='https://huggingface.co/daven3/k2_fp_delta'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging Face-delta%20model-red'></a> <a href='https://huggingface.co/daven3/k2-v1'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging Face-k2%20v1-red'></a> <a href='https://huggingface.co/daven3/k2_it_adapter'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging Face-adapter%20model-red'></img></a> <a href='https://huggingface.co/datasets/daven3/geosignal'><img src='https://img.shields.io/badge/Dataset-GeoSignal-4169E1'></img></a> <a href='https://huggingface.co/datasets/daven3/geobench'><img src='https://img.shields.io/badge/Dataset-GeoBench-4169E1'></img></a>

Introduction

We introduce K2 (7B), an open-source language model trained by firstly further pretraining LLaMA on collected and cleaned geoscience literature, including geoscience open-access papers and Wikipedia pages, and secondly fine-tuning with knowledge-intensive instruction tuning data (GeoSignal). As for preliminary evaluation, we use GeoBench (consisting of NPEE and AP Test on Geology, Geography, and Environmental Science) as the benchmark. K2 outperforms the baselines on objectiv e and subjective tasks compared to several baseline models with similar parameters. In this repository, we will share the following code and data.

The following is the overview of training K2: overview

Quick Start

Installation

1. Prepare the code and the environment

Clone our repository, create a Python environment, and activate it via the following command

git clone https://github.com/davendw49/k2.git
cd k2
conda env create -f k2.yaml
conda activate k2

2. Prepare the pretrained K2 (GeoLLaMA)

The current version of K2 is on Huggingface Model The previous version of K2 consists of two parts: a delta model (like Vicuna), an add-on weight towards LLaMA-7B, and an adapter model (trained via PEFT).

Delta modelAdapter modelFull model
k2_fp_deltak2_it_adapterk2_v1
python -m apply_delta --base /path/to/weights/of/llama --target /path/to/weights/of/geollama/ --delta daven3/k2_fp_delta

Start with Docker

Coming soon...

3. Use K2

base_model = /path/to/k2
tokenizer = LlamaTokenizer.from_pretrained(base_model)
model = LlamaForCausalLM.from_pretrained(
    base_model,
    load_in_8bit=load_8bit,
    device_map=device_map
    torch_dtype=torch.float16
)
model.config.pad_token_id = tokenizer.pad_token_id = 0
model.config.bos_token_id = 1
model.config.eos_token_id = 2

Or, alternatively,

base_model = /path/to/geollama
lora_weights = /path/to/adapter/model
tokenizer = LlamaTokenizer.from_pretrained(base_model)
model = LlamaForCausalLM.from_pretrained(
    base_model,
    load_in_8bit=load_8bit,
    device_map=device_map
    torch_dtype=torch.float16
)
model = PeftModel.from_pretrained(
    model,
    lora_weights,
    torch_dtype=torch.float16,
    device_map=device_map,
)
model.config.pad_token_id = tokenizer.pad_token_id = 0
model.config.bos_token_id = 1
model.config.eos_token_id = 2

Data

In this repo, we share the instruction data and benchmark data:

Further pretrain

Our text corpus for further pretraining on LLaMA-7B consists of 3.9 billion tokens from geoscience papers published in selected high-quality journals in earth science and mainly collected by GAKG.

Delta Model on Hugging Face: daven3/k2_fp_delta

Instruction Tuning: GeoSignal

Scientific domain adaptation has two main steps during instruction tuning.

The following is the illustration of the training domain-specific language model recipe: recipe

Benchmark: GeoBench

In GeoBench, we collect 183 multiple-choice questions in NPEE, and 1,395 in AP Test, for objective tasks. Meanwhile, we gather all 939 subjective questions in NPEE to be the subjective tasks set and use 50 to measure the baselines with human evaluation.

Code

Further Pretrain

The training script is run_clm.py

deepspeed --num_gpus=4 run_clm.py --deepspeed ds_config_zero.json >log 2>&1 &

loss curve

The parameters we use:

- Batch size per device: 2
- Global batch size: 128 (2*4gpu*16gradient accumulation steps) 
- Number of trainable parameters: 6738415616 (7b)
- lr: 1e-5
- bf16: true
- tf32: true
- Warmup: 0.03/3 epoch (nearly 1000 steps)
- Zero_optimization_stage: 3

Tips: We can not resume smoothly from checkpoints for the limited computing power. Therefore, we did not load the optimizer state dict when resuming training. Even though there are two noticeable spikes in the diagram, the performance seems to stay at normal.

Instruction tuning

The training script is finetune.py

python finetune.py --base_model /path/to/checkpoint-30140 --data_path /path/to/alpaca.json --output_dir /path/to/stage/one/model/ --cuda_id 2 --lora_target_modules "q_proj" "k_proj" "v_proj"

python finetune.py --base_model /path/to/checkpoint-30140 --data_path /path/to/geosignal.json --output_dir /path/to/stage/two/model/ --cuda_id 2 --lora_target_modules "q_proj" "k_proj" "v_proj" --resume_from_checkpoint /path/to/stage/one/model/

- batch_size: 128
- micro batch size: 4
- num epochs: 1
- learning rate: 3e-4
- cutoff len: 512
- val set size: 2000
- lora r: 8
- lora alpha: 16
- lora dropout: 0.05
- lora target modules: ["q_proj", "k_proj", "v_proj"]

Cases

Case 1Case 2Case 3

Evaluation

We share the original evaluation code in evaluation folder, and we will release Geo-Eval in the near future with more evaluation methods.

Why named K2 ?

K2 is originally from the name of the second-highest mountain in the world, and we believe that in the future, larger and more powerful geoscience language models will be created. What is more, to train a model to shift to a discipline with a significant domain barrier, we have encountered many difficulties (collecting corpus, clean academic data, computing power, ...), which shares with the fact that climbing K2 is no less challenging than Mount Everest 🏔️.

Contributors

This project was founded by Acemap at Shanghai Jiao Tong University, including Cheng Deng, Tianhang Zhang, Zhongmou He, Qiyuan Chen, Yuanyuan Shi, Le Zhou, supervised by Weinan Zhang, Luoyi Fu, Zhouhan Lin, Junxian He, and Xinbing Wang. The whole project is supported by Chenghu Zhou and the Institute of Geographical Science, Natural Resources Research, Chinese Academy of Sciences.

Acknowledgements

K2 has referred to the following open-source projects. We want to express our gratitude and respect to the researchers of the projects.

K2 is under the support of Chenghu Zhou and the Institute of Geographical Science, Natural Resources Research, Chinese Academy of Sciences.

We would also like to express our appreciation for the effort of data processing from Yutong Xu and Beiya Dai.

License

K2 is a research preview intended for non-commercial use only, subject to the model License of LLaMA and the Terms of Use of the data generated by OpenAI. Please contact us if you find any potential violations. The code is released under the Apache License 2.0. The data GeoSignal and GeoBench is updating occasionally, if you want to subscribe the data, you can emaill us davendw@sjtu.edu.cn.

Citation ArXiv

If you use the code or data of K2, please declare the reference with the following:

@misc{deng2023learning,
      title={K2: A Foundation Language Model for Geoscience Knowledge Understanding and Utilization}, 
      author={Cheng Deng and Tianhang Zhang and Zhongmou He and Yi Xu and Qiyuan Chen and Yuanyuan Shi and Luoyi Fu and Weinan Zhang and Xinbing Wang and Chenghu Zhou and Zhouhan Lin and Junxian He},
      year={2023},
      eprint={2306.05064},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}