README.md

---
license: apache-2.0
pipeline_tag: text-generation
---




<p align="center">
    <b><font size="6">SongComposer</font></b> 
<p>

<div align="center">

[💻Github Repo](https://github.com/pjlab-songcomposer/songcomposer)

[📖Paper](https://arxiv.org/abs/2402.17645)

</div>

**SongComposer** is a language large model (LLM) based on [InternLM2](https://github.com/InternLM/InternLM) for lyric and melody composition in song generation. 

We release SongComposer series in two versions:

- SongComposer_pretrain: The pretrained SongComposer with InternLM2 as the initialization of the LLM, gains basic knowledge of lyric and melody.
- SongComposer_sft: The finetuned SongComposer for *instruction-following song generation* including lyric to melody, melody to lyric, song continuation, text to song.

### Import from Transformers
To load the SongComposer_pretrain model using Transformers, use the following code:
```python
from transformers import AutoTokenizer, AutoModel
ckpt_path = "Mar2Ding/songcomposer_pretrain"
tokenizer = AutoTokenizer.from_pretrained(ckpt_path, trust_remote_code=True)
model = AutoModel.from_pretrained(ckpt_path, trust_remote_code=True).cuda().half()
prompt = '<bop> Total 7 lines. The first line:可,<D4>,<137>,<79>|惜,<D#4>,<137>,<79>|这,<F4>,<137>,<88>|是,<F4>,<121>,<79>|属,<F4>,<121>,<79>|于,<D#4>,<214>,<88>|你,<D#4>,<141>,<79>|的,<D4>,<130>,<79>|风,<C4>,<151>,<79>|景,<A#3> <F3>,<181><137>,<79>\n'
model.inference_pretrain(prompt, tokenizer)
```

### 通过 Transformers 加载
通过以下的代码加载 SongComposer_pretrain 模型
```python
from transformers import AutoTokenizer, AutoModel
ckpt_path = "Mar2Ding/songcomposer_pretrain"
tokenizer = AutoTokenizer.from_pretrained(ckpt_path, trust_remote_code=True)
model = AutoModel.from_pretrained(ckpt_path, trust_remote_code=True).cuda().half()
prompt = '<bop> Total 7 lines. The first line:可,<D4>,<137>,<79>|惜,<D#4>,<137>,<79>|这,<F4>,<137>,<88>|是,<F4>,<121>,<79>|属,<F4>,<121>,<79>|于,<D#4>,<214>,<88>|你,<D#4>,<141>,<79>|的,<D4>,<130>,<79>|风,<C4>,<151>,<79>|景,<A#3> <F3>,<181><137>,<79>\n'
model.inference_pretrain(prompt, tokenizer)
```

### Open Source License
The code is licensed under Apache-2.0, while model weights are fully open for academic research and also allow free commercial usage.