Important Notice: This is a pre-trained base model without instruction-following capabilities.

YuLan-Mini is a lightweight language model with 2.4 billion parameters. It achieves performance comparable to industry-leading models trained on significantly more data, despite being pre-trained on only 1.08T tokens. The model excels particularly in the domains of mathematics and code. To facilitate reproducibility, we will open-source the relevant pre-training resources.

---

Model Downloads 🔗

---

Features 🌟

Our pre-training methodology improves training efficiency through three key innovations:

1. an elaborately designed data pipeline that combines data cleaning with data schedule strategies;
2. a systematic optimization method that can effectively mitigate training instability;
3. an effective annealing approach that integrate targeted data selection and long context training.

---

Behchmarks 🌟

---

Pre-Training Resources 🔧

To enhance research transparency and reproducibility, we are open-sourcing relevant pre-training resources:

1. Pre-training and Evaluation Code

The pre-training and evaluation code will be released in a future update.

2. Intermediate Stage Checkpoints

The intermediate stage checkpoints are released in YuLan-Mini.

Stage	Curriculum Phase	4K Context	28K Context	Optimizer	Inference Architecture	LAMBADA Acc	GSM8K Acc	HumanEval pass@1
Stable	5	YuLan-Mini-Phase5			yulanmini	53.85	3.41	12.26
Stable	10	YuLan-Mini-Phase10			yulanmini	55.00	9.57	15.95
Stable	15	YuLan-Mini-Phase15		✅	yulanmini	55.81	13.81	16.99
Stable	20	YuLan-Mini-Phase20		✅	yulanmini	55.81	21.39	20.79
Stable	25 (1T tokens)	YuLan-Mini-Before-Annealing		✅	yulanmini	55.67	29.94	34.06
Annealing	26	YuLan-Mini-4K			llama*	64.72	66.65	61.60
Annealing	27		YuLan-Mini		llama*	65.67	68.46	64.00

*: For easier inference and deployment, we merged the re-parameterized added parameters and scaling factors into the final released models (YuLan-Mini and YuLan-Mini-Intermediate-4K), enabling it to run on the Llama architecture. However, these parameters are still retained in the intermediate checkpoints from the training process.

3. Optimizer States Before Annealing

YuLan-Mini-Before-Annealing

4. The Used Open-Source Datasets

Used-Datasets-List

5. Data Distribution for every phase

6. Synthetic Data

Data cleaning and synthesis pipeline:

The synthetic data we are using is released in YuLan-Mini-Datasets

What you can do with these pre-training resources

1. Pre-train your own LLM. You can use our data and curriculum to train a model that's just as powerful as YuLan-Mini.
2. Perform your own learning rate annealing. During the annealing phase, YuLan-Mini's learning ability is at its peak. You can resume training from the checkpoint before annealing and use your own dataset for learning rate annealing.
3. Fine-tune the Instruct version of the LLM. You can use the YuLan-Mini base model to train your own Instruct version.
4. Training dynamics research. You can use YuLan-Mini's intermediate checkpoints to explore internal changes during the pre-training process.
5. Synthesize your own data. You can use YuLan-Mini's data pipeline to clean and generate your own dataset.

---

Quick Start 💻

Below is a simple example for inference using Huggingface:

Huggingface Inference Example

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("yulan-team/YuLan-Mini")
model = AutoModelForCausalLM.from_pretrained("yulan-team/YuLan-Mini", torch_dtype=torch.bfloat16)

# Input text
input_text = "Renmin University of China is"
inputs = tokenizer(input_text, return_tensors="pt")

# Completion
output = model.generate(inputs["input_ids"], max_new_tokens=100)
print(tokenizer.decode(output[0], skip_special_tokens=True))

vLLM Serve Example

vllm serve yulan-team/YuLan-Mini --dtype bfloat16

SGLang Serve Example

python -m sglang.launch_server --model-path yulan-team/YuLan-Mini --port 30000 --host 0.0.0.0

---

The Team

YuLan-Mini is developed and maintained by AI Box, Renmin University of China.

License

• The code in this repository is released under the MIT License.
• Policies regarding the use of model weights, intermediate optimizer states, and training data will be announced in future updates.
• Limitations: Despite our efforts to mitigate safety concerns and encourage the generation of ethical and lawful text, the probabilistic nature of language models may still lead to unexpected outputs. For instance, responses might contain bias, discrimination, or other harmful content. Please refrain from disseminating such content. We are not liable for any consequences arising from the spread of harmful information.

Citation

If you find YuLan-Mini helpful for your research or development, please cite our technical report:

@misc{hu2024yulanmini,
      title={YuLan-Mini: An Open Data-efficient Language Model}, 
      author={Yiwen Hu and Huatong Song and Jia Deng and Jiapeng Wang and Jie Chen and Kun Zhou and Yutao Zhu and Jinhao Jiang and Zican Dong and Wayne Xin Zhao and Ji-Rong Wen},
      year={2024},
      eprint={2412.17743},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.17743}, 
}