license: mit
language:
- en
metrics:
- accuracy
pipeline_tag: text-generation
🎼 ChatMusician: Fostering Intrinsic Musical Abilities Into LLM
🌐 DemoPage | 🤗 Dataset | 🤗 Benchmark | 📖 arXiv | Code
🔔News
- 🔥[2023-12-10]: The release of ChatMusician's demo, code, model, data, and benchmark. 😆
- [2023-11-30]: Checkout another awesome project MMMU that includes multimodal music reasoning.
Introduction
While Large Language Models (LLMs) demonstrate impressive capabilities in text generation, we find that their ability has yet to be generalized to music, humanity’s creative language. We introduce ChatMusician, an open-source LLM that integrates intrinsic musical abilities.
It is based on continual pre-training and finetuning LLaMA2 on a text-compatible music representation, ABC notation, and the music is treated as a second language. ChatMusician can understand and generate music with a pure text tokenizer without any external multi-modal neural structures or tokenizers. Interestingly, endowing musical abilities does not harm language abilities, even achieving a slightly higher MMLU score. Our model is capable of composing well-structured, full-length music, conditioned on texts, chords, melodies, motifs, musical forms, etc, surpassing GPT-4 baseline. On our meticulously curated college-level music understanding benchmark, MusicTheoryBench, ChatMusician surpasses LLaMA2 and GPT-3.5 on zero-shot setting by a noticeable margin. Our work reveals that LLMs can be an excellent compressor for music, but there remains significant territory to be conquered. Code, data, model, and benchmark are open-sourced.
ChatMusician-Base is a pretrained model. ChatMusician is recommended for producing symbolic music.
Training Data
ChatMusician-Base is pretrained on the 🤗 MusicPile, which is the first pretraining corpus for developing musical abilities in large language models. Check out the dataset card for more details.
Training Procedure
We initialized a fp16-precision ChatMusician-Base from the LLaMA2-7B-Base weights, and applied a continual pre-training plus fine-tuning pipeline. LoRA adapters were integrated into the attention and MLP layers, with additional training on embeddings and all linear layers. The maximum sequence length was 2048. We utilized 16 80GB-A800 GPUs for one epoch pre-training. DeepSpeed was employed for memory efficiency, and the AdamW optimizer was used with a 1e-4 learning rate and a 5% warmup cosine scheduler. Gradient clipping was set at 1.0. The LoRA parameters dimension, alpha, and dropout were set to 64, 16, and 0.1, with a batch size of 8.
Intended Uses
These models are trained for research purposes. They are designed to solve general math problems. They can be used in educational software, tutoring systems, or any application where a solution to a math problem is needed. The models can generate both a chain of thought (CoT) rationale and a program of thought (PoT) rationale, providing a comprehensive solution to a given math problem.
Limitations
We've tried our best to build math generalist models. However, we acknowledge that the models' performance may vary based on the complexity and specifics of the math problem. Still not all mathematical fields can be covered comprehensively.
Citation
If you find our work helpful, feel free to give us a cite.
@misc{yuan2024chatmusician,
title={ChatMusician: Understanding and Generating Music Intrinsically with LLM},
author={Ruibin Yuan and Hanfeng Lin and Yi Wang and Zeyue Tian and Shangda Wu and Tianhao Shen and Ge Zhang and Yuhang Wu and Cong Liu and Ziya Zhou and Ziyang Ma and Liumeng Xue and Ziyu Wang and Qin Liu and Tianyu Zheng and Yizhi Li and Yinghao Ma and Yiming Liang and Xiaowei Chi and Ruibo Liu and Zili Wang and Pengfei Li and Jingcheng Wu and Chenghua Lin and Qifeng Liu and Tao Jiang and Wenhao Huang and Wenhu Chen and Emmanouil Benetos and Jie Fu and Gus Xia and Roger Dannenberg and Wei Xue and Shiyin Kang and Yike Guo},
year={2024},
eprint={2402.16153},
archivePrefix={arXiv},
primaryClass={cs.SD}
}