JiuZhou: Open Foundation Language Models for Geoscience

🎉 News

[2025-05] Paper TagRouter: Learning Route to LLMs through Tags for Open-Domain Text Generation Tasks has been accepted by the top NLP conference ACL. Model Download.
[2025-03] Paper GeoFactory: an LLM Performance Enhancement Framework for Geoscience Factual and Inferential Tasks has been accepted by the journal Big Earth Data. Data Download.
[2025-03] Paper ClimateChat: Designing Data and Methods for Instruction Tuning LLMs to Answer Climate Change Queries has been accepted by the International Conference on Learning Representations (ICLR). Model Download.
[2024-12] Paper JiuZhou: Open Foundation Language Models and Effective Pre-training Framework for Geoscience has been accepted by the International Journal of Digital Earth. Model Introduction. Project Repository.
[2024-09] Released chat model ClimateChat.
[2024-08] Paper PreparedLLM: Effective Pre-pretraining Framework for Domain-specific Large Language Models has been accepted by the journal Big Earth Data. WeChat article: PreparedLLM: Effective Pre-pretraining Framework for Domain-specific Large Language Models. Model Download.
[2024-08] Released chat model Chinese-Mistral-7B-Instruct-v0.2, featuring significantly improved language understanding and multi-turn conversation capabilities.
[2024-06] Released chat model JiuZhou-Instruct-v0.2, with significantly enhanced language understanding and multi-turn conversation capabilities.
[2024-05] WeChat Article: Chinese Vocabulary Expansion Incremental Pretraining for Large Language Models: Chinese-Mistral Released.
[2024-03] Released base model Chinese-Mistral-7B-v0.1 and chat model Chinese-Mistral-7B-Instruct-v0.1. Model Introduction. Project Repository.
[2024-03] Released JiuZhou's base version JiuZhou-base, instruct version JiuZhou-instruct-v0.1, and intermediate checkpoints. Model Introduction. Project Repository.
[2024-01] Completed training of Chinese-Mistral and JiuZhou, and commenced model evaluation.

Introduction
Download
Inference
Model Performance
Model Training Process
Model Training Code
Citations
Acknowledgments

Introduction

The field of geoscience has amassed a vast amount of data, necessitating the extraction and integration of diverse knowledge from this data to address global change challenges, promote sustainable development, and accelerate scientific discovery. Foundation language models initially learn and integrate knowledge autonomously through self-supervised pre-training on extensive text data. Subsequently, they acquire the capability to solve geoscience problems through instruction tuning. However, when the foundational language models lack sufficient geoscience expertise, instruction tuning with relevant data can lead to the generation of content that is inconsistent with established facts. To improve the model's accuracy and practicality, a robust geoscience foundational language model is urgently needed.

This study uses Mistral-7B-v0.1 as the base model and continues pretraining on a large geoscience corpus. It also incorporates the domain-specific large language model pre-pretraining framework (PreparedLLM) and the "two-stage pre-adaptation pre-training" algorithm to build the geoscience large language model, JiuZhou.

Download

Model Series	Model	Download Link	Description
JiuZhou	JiuZhou-base	Huggingface	Base model (Rich in geoscience knowledge)
JiuZhou	JiuZhou-Instruct-v0.1	Huggingface	Instruct model (Instruction alignment caused a loss of some geoscience knowledge, but it has instruction-following ability) LoRA fine-tuned on Alpaca_GPT4 in both Chinese and English and GeoSignal
JiuZhou	JiuZhou-Instruct-v0.2	HuggingFace Wisemodel	Instruct model (Instruction alignment caused a loss of some geoscience knowledge, but it has instruction-following ability) Fine-tuned with high-quality general instruction data
ClimateChat	ClimateChat	HuggingFace Wisemodel	Instruct model Fine-tuned on JiuZhou-base for instruction following
Chinese-Mistral	Chinese-Mistral-7B	HuggingFace Wisemodel ModelScope	Base model
Chinese-Mistral	Chinese-Mistral-7B-Instruct-v0.1	HuggingFace Wisemodel ModelScope	Instruct model LoRA fine-tuned with Alpaca_GPT4 in both Chinese and English
Chinese-Mistral	Chinese-Mistral-7B-Instruct-v0.2	HuggingFace Wisemodel	Instruct model LoRA fine-tuned with a million high-quality instructions
PreparedLLM	Prepared-Llama	Huggingface Wisemodel	Base model Continual pretraining with a small number of geoscience data Recommended to use JiuZhou

Inference

Below is an example of inference code using JiuZhou-Instruct-v0.2.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")

model_path = "itpossible/JiuZhou-Instruct-v0.2"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, device_map=device)

text = "What is geoscience?"
messages = [{"role": "user", "content": text}]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(device)
outputs_id = model.generate(inputs, max_new_tokens=600, do_sample=True)
outputs = tokenizer.batch_decode(outputs_id, skip_special_tokens=True)[0]
print(outputs)

Model Performance

Geoscience Ability

We evaluate the performance of JiuZhou using the GeoBench benchmark.
JiuZhou outperforms GPT-3.5 in objective tasks:

JiuZhou also scores higher than baselines across six criteria in subjective tasks:

### General Ability We evaluate the performance of JiuZhou using three benchmark datasets: C-Eval, CMMLU, and MMLU.
Compared to other variants of Llama and Mistral models, JiuZhou shows outstanding performance:

## Model Training Process ### Training Corpus The corpus consists of 50 million general documents and 3.4 million geoscience-related documents.

### Training Framework We use the JiuZhou-Framework proposed in this study.

### Two-stage Pre-adaptation Pre-training (TSPT) TSPT improves the efficiency of using limited geoscience data and overcomes some of the technical bottlenecks in continual pretraining for LLMs.
The difference between TSPT and single-stage training algorithms:

Comparison of TSPT and one-stage pre-training algorithm performance:

## Model Training Code We use [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) to fine-tune JiuZhou.

Project Deployment

git clone https://github.com/THU-ESIS/JiuZhou.git
cd JiuZhou
pip install -e ".[torch,metrics]"

Model Training

Pre-training：

llamafactory-cli train examples/train_lora/JiuZhou_pretrain_sft.yaml

Instruction-tuning：

llamafactory-cli train examples/train_lora/JiuZhou_lora_sft.yaml

Chat with the fine-tuned JiuZhou:：

llamafactory-cli chat examples/inference/JiuZhou_lora_sft.yaml

Merge the instruction-tuned LoRA weights with the original JiuZhou weights:

llamafactory-cli export examples/merge_lora/JiuZhou_lora_sft.yaml

Citations

@article{chen2024preparedllm,
  author = {Chen, Zhou and Lin, Ming and Wang, Zimeng and Zang, Mingrun and Bai, Yuqi},
  title = {PreparedLLM: Effective Pre-pretraining Framework for Domain-specific Large Language Models},
  year = {2024},
  journal = {Big Earth Data},
  pages = {1--24},
  doi = {10.1080/20964471.2024.2396159},
  url = {https://doi.org/10.1080/20964471.2024.2396159}
}

itpossible
/

JiuZhou-base