Update README.md
Browse files
README.md
CHANGED
@@ -25,6 +25,57 @@ margin. Our work reveals that LLMs can be an excellent compressor for music, but
|
|
25 |
|
26 |
<!-- <audio controls src="https://cdn-uploads.huggingface.co/production/uploads/5fd6f670053c8345eddc1b68/8NSONUjIF7KGUCfwzPCd9.mpga"></audio> -->
|
27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
## Example Prompts
|
29 |
|
30 |
**Our model produces symbolic music(ABC notation) well in the following prompts.** Here are some musical tasks.
|
@@ -144,56 +195,7 @@ We included GPT-3.5, GPT-4, LLaMA2-7B-Base, ChatMusician-Base, and ChatMusician.
|
|
144 |
2. General language abilities of ChatMusician are evaluated on the [Massive Multitask Language Understanding (MMLU) dataset](https://huggingface.co/datasets/lukaemon/mmlu).
|
145 |
|
146 |
|
147 |
-
## Usage
|
148 |
-
|
149 |
-
You can use the models through Huggingface's Transformers library. Check our Github repo for more advanced use: [https://github.com/hf-lin/ChatMusician](https://github.com/hf-lin/ChatMusician)
|
150 |
-
|
151 |
-
## CLI demo
|
152 |
-
```python
|
153 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
|
154 |
-
from string import Template
|
155 |
-
prompt_template = Template("Human: ${inst} </s> Assistant: ")
|
156 |
-
|
157 |
-
tokenizer = AutoTokenizer.from_pretrained("m-a-p/ChatMusician-v1-sft-78k", trust_remote_code=True)
|
158 |
-
model = AutoModelForCausalLM.from_pretrained("m-a-p/ChatMusician-v1-sft-78k", trust_remote_code=True).eval()
|
159 |
-
model.cuda()
|
160 |
-
generation_config = GenerationConfig(
|
161 |
-
temperature=0.2,
|
162 |
-
top_k=40,
|
163 |
-
top_p=0.9,
|
164 |
-
do_sample=True,
|
165 |
-
num_beams=1,
|
166 |
-
repetition_penalty=1.1,
|
167 |
-
min_new_tokens=10,
|
168 |
-
max_new_tokens=1536
|
169 |
-
)
|
170 |
-
|
171 |
-
instruction = """Using ABC notation, recreate the given text as a musical score.
|
172 |
-
Meter C
|
173 |
-
Notes The parts are commonly interchanged.
|
174 |
-
Transcription 1997 by John Chambers
|
175 |
-
Key D
|
176 |
-
Note Length 1/8
|
177 |
-
Rhythm reel
|
178 |
-
"""
|
179 |
-
|
180 |
-
prompt = prompt_template.safe_substitute({"inst": instruction})
|
181 |
-
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
|
182 |
-
response = model.generate(
|
183 |
-
input_ids=inputs["input_ids"].to(model.device),
|
184 |
-
attention_mask=inputs['attention_mask'].to(model.device),
|
185 |
-
eos_token_id=tokenizer.eos_token_id,
|
186 |
-
generation_config=generation_config,
|
187 |
-
)
|
188 |
-
response = tokenizer.decode(response[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
|
189 |
-
print(response)
|
190 |
-
```
|
191 |
|
192 |
-
## Limitations
|
193 |
-
- For the music tasks, the model currently only supports strict format and close-ended instructions.
|
194 |
-
- The model suffers from hallucinations, and shouldn't be used for music education.
|
195 |
-
- A large proportion of the training data is in the style of Irish music.
|
196 |
-
- The MusicThoeryBench results reported in the paper are obtained with perplexity mode.
|
197 |
|
198 |
## Citation
|
199 |
If you find our work helpful, feel free to give us a cite.
|
|
|
25 |
|
26 |
<!-- <audio controls src="https://cdn-uploads.huggingface.co/production/uploads/5fd6f670053c8345eddc1b68/8NSONUjIF7KGUCfwzPCd9.mpga"></audio> -->
|
27 |
|
28 |
+
## Usage
|
29 |
+
|
30 |
+
You can use the models through Huggingface's Transformers library. Check our Github repo for more advanced use: [https://github.com/hf-lin/ChatMusician](https://github.com/hf-lin/ChatMusician)
|
31 |
+
|
32 |
+
## CLI demo
|
33 |
+
```python
|
34 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
|
35 |
+
from string import Template
|
36 |
+
prompt_template = Template("Human: ${inst} </s> Assistant: ")
|
37 |
+
|
38 |
+
tokenizer = AutoTokenizer.from_pretrained("m-a-p/ChatMusician-v1-sft-78k", trust_remote_code=True)
|
39 |
+
model = AutoModelForCausalLM.from_pretrained("m-a-p/ChatMusician-v1-sft-78k", trust_remote_code=True).eval()
|
40 |
+
model.cuda()
|
41 |
+
generation_config = GenerationConfig(
|
42 |
+
temperature=0.2,
|
43 |
+
top_k=40,
|
44 |
+
top_p=0.9,
|
45 |
+
do_sample=True,
|
46 |
+
num_beams=1,
|
47 |
+
repetition_penalty=1.1,
|
48 |
+
min_new_tokens=10,
|
49 |
+
max_new_tokens=1536
|
50 |
+
)
|
51 |
+
|
52 |
+
instruction = """Using ABC notation, recreate the given text as a musical score.
|
53 |
+
Meter C
|
54 |
+
Notes The parts are commonly interchanged.
|
55 |
+
Transcription 1997 by John Chambers
|
56 |
+
Key D
|
57 |
+
Note Length 1/8
|
58 |
+
Rhythm reel
|
59 |
+
"""
|
60 |
+
|
61 |
+
prompt = prompt_template.safe_substitute({"inst": instruction})
|
62 |
+
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
|
63 |
+
response = model.generate(
|
64 |
+
input_ids=inputs["input_ids"].to(model.device),
|
65 |
+
attention_mask=inputs['attention_mask'].to(model.device),
|
66 |
+
eos_token_id=tokenizer.eos_token_id,
|
67 |
+
generation_config=generation_config,
|
68 |
+
)
|
69 |
+
response = tokenizer.decode(response[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
|
70 |
+
print(response)
|
71 |
+
```
|
72 |
+
|
73 |
+
## Limitations
|
74 |
+
- For the music tasks, the model currently only supports strict format and close-ended instructions.
|
75 |
+
- The model suffers from hallucinations, and shouldn't be used for music education.
|
76 |
+
- A large proportion of the training data is in the style of Irish music.
|
77 |
+
- The MusicThoeryBench results reported in the paper are obtained with perplexity mode.
|
78 |
+
|
79 |
## Example Prompts
|
80 |
|
81 |
**Our model produces symbolic music(ABC notation) well in the following prompts.** Here are some musical tasks.
|
|
|
195 |
2. General language abilities of ChatMusician are evaluated on the [Massive Multitask Language Understanding (MMLU) dataset](https://huggingface.co/datasets/lukaemon/mmlu).
|
196 |
|
197 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
198 |
|
|
|
|
|
|
|
|
|
|
|
199 |
|
200 |
## Citation
|
201 |
If you find our work helpful, feel free to give us a cite.
|