Commit
·
abb8ae6
1
Parent(s):
fb62397
Update README.md
Browse filesUpdated README with chat_template
README.md
CHANGED
|
@@ -38,6 +38,15 @@ DeciLM-7B-instruct is a derivative of the recently released [DeciLM-7B](https://
|
|
| 38 |
- **Finetuning Notebook:** [DeciLM-7B Finetuning Notebook](https://colab.research.google.com/drive/1kEV6i96AQ94xTCvSd11TxkEaksTb5o3U?usp=sharing)
|
| 39 |
- **Text Generation Notebook:** [DeciLM-7B-instruct Text Generation Notebook](https://bit.ly/declm-7b-instruct)
|
| 40 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
## Uses
|
| 42 |
|
| 43 |
The model is intended for commercial and research use in English.
|
|
@@ -54,16 +63,21 @@ model_name = "Deci/DeciLM-7B-instruct"
|
|
| 54 |
|
| 55 |
device = "cuda" # for GPU usage or "cpu" for CPU usage
|
| 56 |
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
|
| 62 |
model = AutoModelForCausalLM.from_pretrained(
|
| 63 |
model_name,
|
| 64 |
device_map="auto",
|
| 65 |
trust_remote_code=True,
|
| 66 |
-
|
| 67 |
)
|
| 68 |
|
| 69 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
@@ -75,25 +89,19 @@ deci_generator = pipeline("text-generation",
|
|
| 75 |
temperature=0.1,
|
| 76 |
device_map="auto",
|
| 77 |
max_length=4096,
|
| 78 |
-
return_full_text=False
|
| 79 |
-
)
|
| 80 |
|
| 81 |
-
|
| 82 |
|
| 83 |
-
|
| 84 |
-
### System:
|
| 85 |
-
You are an AI assistant that follows instruction extremely well. Help as much as you can.
|
| 86 |
-
### User:
|
| 87 |
-
{instruction}
|
| 88 |
-
### Assistant:
|
| 89 |
-
"""
|
| 90 |
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
|
|
|
| 94 |
|
| 95 |
-
response = deci_generator(
|
| 96 |
-
print(response)
|
| 97 |
```
|
| 98 |
|
| 99 |
## Evaluation
|
|
|
|
| 38 |
- **Finetuning Notebook:** [DeciLM-7B Finetuning Notebook](https://colab.research.google.com/drive/1kEV6i96AQ94xTCvSd11TxkEaksTb5o3U?usp=sharing)
|
| 39 |
- **Text Generation Notebook:** [DeciLM-7B-instruct Text Generation Notebook](https://bit.ly/declm-7b-instruct)
|
| 40 |
|
| 41 |
+
### Prompt Template
|
| 42 |
+
```
|
| 43 |
+
### System:
|
| 44 |
+
{system_prompt}
|
| 45 |
+
### User:
|
| 46 |
+
{user_prompt}
|
| 47 |
+
### Assistant:
|
| 48 |
+
```
|
| 49 |
+
|
| 50 |
## Uses
|
| 51 |
|
| 52 |
The model is intended for commercial and research use in English.
|
|
|
|
| 63 |
|
| 64 |
device = "cuda" # for GPU usage or "cpu" for CPU usage
|
| 65 |
|
| 66 |
+
quantize = False # Optional. Useful for GPUs with less than 24GB memory
|
| 67 |
+
|
| 68 |
+
if quantize:
|
| 69 |
+
dtype_kwargs = dict(quantization_config=BitsAndBytesConfig(
|
| 70 |
+
load_in_4bit=True,
|
| 71 |
+
bnb_4bit_compute_dtype=torch.bfloat16
|
| 72 |
+
))
|
| 73 |
+
else:
|
| 74 |
+
dtype_kwargs = dict(torch_dtype="auto")
|
| 75 |
|
| 76 |
model = AutoModelForCausalLM.from_pretrained(
|
| 77 |
model_name,
|
| 78 |
device_map="auto",
|
| 79 |
trust_remote_code=True,
|
| 80 |
+
**dtype_kwargs
|
| 81 |
)
|
| 82 |
|
| 83 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
|
|
| 89 |
temperature=0.1,
|
| 90 |
device_map="auto",
|
| 91 |
max_length=4096,
|
| 92 |
+
return_full_text=False)
|
|
|
|
| 93 |
|
| 94 |
+
system_prompt = "You are an AI assistant that follows instruction extremely well. Help as much as you can."
|
| 95 |
|
| 96 |
+
user_prompt = "How do I make the most delicious pancakes the world has ever tasted?"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 97 |
|
| 98 |
+
prompt = tokenizer.apply_chat_template([
|
| 99 |
+
{"role": "system", "content": system_prompt},
|
| 100 |
+
{"role": "user", "content": user_prompt},
|
| 101 |
+
], tokenize=False, add_generation_prompt=True)
|
| 102 |
|
| 103 |
+
response = deci_generator(prompt)[0]['generated_text']
|
| 104 |
+
print(prompt + response)
|
| 105 |
```
|
| 106 |
|
| 107 |
## Evaluation
|