mattdangerw
commited on
Commit
•
ab2afa8
1
Parent(s):
be70570
Update README.md with new model card content
Browse files
README.md
CHANGED
@@ -1,23 +1,170 @@
|
|
1 |
---
|
2 |
library_name: keras-hub
|
3 |
---
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
*
|
12 |
-
*
|
13 |
-
*
|
14 |
-
*
|
15 |
-
*
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
library_name: keras-hub
|
3 |
---
|
4 |
+
### Model Overview
|
5 |
+
Phi-3 is a set of large language models published by Microsoft. Models are instruction tuned, and range in size from 3 billion to 14 billion parameters. See the model card below for benchmarks, data sources, and intended use cases.
|
6 |
+
|
7 |
+
Weights are released under the [MIT License](https://opensource.org/license/mit). Keras model code is released under the [Apache 2 License](https://github.com/keras-team/keras-hub/blob/master/LICENSE).
|
8 |
+
|
9 |
+
## Links
|
10 |
+
|
11 |
+
* [Phi-3 Quickstart Notebook](https://www.kaggle.com/code/matthewdwatson/phi-3-quickstart)
|
12 |
+
* [Phi-3 API Documentation](https://keras.io/api/keras_hub/models/phi3/)
|
13 |
+
* [Phi-3 Model Card](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct)
|
14 |
+
* [KerasHub Beginner Guide](https://keras.io/guides/keras_hub/getting_started/)
|
15 |
+
* [KerasHub Model Publishing Guide](https://keras.io/guides/keras_hub/upload/)
|
16 |
+
|
17 |
+
## Installation
|
18 |
+
|
19 |
+
Keras and KerasHub can be installed with:
|
20 |
+
|
21 |
+
```
|
22 |
+
pip install -U -q keras-hub
|
23 |
+
pip install -U -q keras>=3
|
24 |
+
```
|
25 |
+
|
26 |
+
Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instruction on installing them in another environment see the [Keras Getting Started](https://keras.io/getting_started/) page.
|
27 |
+
|
28 |
+
## Presets
|
29 |
+
|
30 |
+
The following model checkpoints are provided by the Keras team. Full code examples for each are available below.
|
31 |
+
|
32 |
+
| Preset name | Parameters | Description |
|
33 |
+
|-----------------------|------------|---------------|
|
34 |
+
|` phi3_mini_4k_instruct_en` | 3.82B | 3B model with 4K max context |
|
35 |
+
| `phi3_mini_128k_instruct_en` | 3.82B | 3B model with 128K max context |
|
36 |
+
|
37 |
+
## Prompts
|
38 |
+
|
39 |
+
Phi-3 models are instruction tuned on turn by turn conversations and should be prompted with examples that precisely match the training data. Specifically, you must alternate user and assistant turns that begin and end with special tokens. New lines do matter. See the following for an example:
|
40 |
+
|
41 |
+
```python
|
42 |
+
prompt = """<|user|>
|
43 |
+
Hello!<|end|>
|
44 |
+
<|assistant|>
|
45 |
+
Hello! How are you?<|end|>
|
46 |
+
<|user|>
|
47 |
+
I'm great. Could you help me with a task?<|end|>
|
48 |
+
"""
|
49 |
+
```
|
50 |
+
|
51 |
+
## Example Usage
|
52 |
+
```shell
|
53 |
+
pip install -U -q keras-hub
|
54 |
+
```
|
55 |
+
|
56 |
+
```python
|
57 |
+
import keras
|
58 |
+
import keras_hub
|
59 |
+
import numpy as np
|
60 |
+
```
|
61 |
+
|
62 |
+
Use `generate()` to do text generation.
|
63 |
+
|
64 |
+
```python
|
65 |
+
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("phi3_mini_128k_instruct_en")
|
66 |
+
phi3_lm.generate("<|user|>\nHow to explain Internet for a medieval knight?<|end|>\n<|assistant|>", max_length=500)
|
67 |
+
|
68 |
+
# Generate with batched prompts.
|
69 |
+
phi3_lm.generate([
|
70 |
+
"<|user|>\nWhat is Keras?<|end|>\n<|assistant|>",
|
71 |
+
"<|user|>\nGive me your best brownie recipe.<|end|>\n<|assistant|>",
|
72 |
+
], max_length=500)
|
73 |
+
```
|
74 |
+
|
75 |
+
Compile the `generate()` function with a custom sampler.
|
76 |
+
|
77 |
+
```python
|
78 |
+
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("phi3_mini_128k_instruct_en")
|
79 |
+
phi3_lm.compile(sampler="greedy")
|
80 |
+
phi3_lm.generate("<|user|>\nWhat is Keras?<|end|>\n<|assistant|>", max_length=30)
|
81 |
+
|
82 |
+
phi3_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
|
83 |
+
phi3_lm.generate("<|user|>\nWhat is Keras?<|end|>\n<|assistant|>", max_length=30)
|
84 |
+
```
|
85 |
+
|
86 |
+
Use `generate()` without preprocessing.
|
87 |
+
|
88 |
+
```python
|
89 |
+
prompt = {
|
90 |
+
"token_ids": np.array([[306, 864, 304, 1827, 0, 0, 0, 0, 0, 0]] * 2),
|
91 |
+
# Use `"padding_mask"` to indicate values that should not be overridden.
|
92 |
+
"padding_mask": np.array([[1, 1, 1, 1, 0, 0, 0, 0, 0, 0]] * 2),
|
93 |
+
}
|
94 |
+
|
95 |
+
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset(
|
96 |
+
"phi3_mini_128k_instruct_en",
|
97 |
+
preprocessor=None,
|
98 |
+
dtype="bfloat16"
|
99 |
+
)
|
100 |
+
phi3_lm.generate(prompt)
|
101 |
+
```
|
102 |
+
|
103 |
+
Call `fit()` on a single batch.
|
104 |
+
|
105 |
+
```python
|
106 |
+
features = ["The quick brown fox jumped.", "I forgot my homework."]
|
107 |
+
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("phi3_mini_128k_instruct_en")
|
108 |
+
phi3_lm.fit(x=features, batch_size=2)
|
109 |
+
```
|
110 |
+
|
111 |
+
## Example Usage with Hugging Face URI
|
112 |
+
|
113 |
+
```shell
|
114 |
+
pip install -U -q keras-hub
|
115 |
+
```
|
116 |
+
|
117 |
+
```python
|
118 |
+
import keras
|
119 |
+
import keras_hub
|
120 |
+
import numpy as np
|
121 |
+
```
|
122 |
+
|
123 |
+
Use `generate()` to do text generation.
|
124 |
+
|
125 |
+
```python
|
126 |
+
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("hf://keras/phi3_mini_128k_instruct_en")
|
127 |
+
phi3_lm.generate("<|user|>\nHow to explain Internet for a medieval knight?<|end|>\n<|assistant|>", max_length=500)
|
128 |
+
|
129 |
+
# Generate with batched prompts.
|
130 |
+
phi3_lm.generate([
|
131 |
+
"<|user|>\nWhat is Keras?<|end|>\n<|assistant|>",
|
132 |
+
"<|user|>\nGive me your best brownie recipe.<|end|>\n<|assistant|>",
|
133 |
+
], max_length=500)
|
134 |
+
```
|
135 |
+
|
136 |
+
Compile the `generate()` function with a custom sampler.
|
137 |
+
|
138 |
+
```python
|
139 |
+
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("hf://keras/phi3_mini_128k_instruct_en")
|
140 |
+
phi3_lm.compile(sampler="greedy")
|
141 |
+
phi3_lm.generate("<|user|>\nWhat is Keras?<|end|>\n<|assistant|>", max_length=30)
|
142 |
+
|
143 |
+
phi3_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
|
144 |
+
phi3_lm.generate("<|user|>\nWhat is Keras?<|end|>\n<|assistant|>", max_length=30)
|
145 |
+
```
|
146 |
+
|
147 |
+
Use `generate()` without preprocessing.
|
148 |
+
|
149 |
+
```python
|
150 |
+
prompt = {
|
151 |
+
"token_ids": np.array([[306, 864, 304, 1827, 0, 0, 0, 0, 0, 0]] * 2),
|
152 |
+
# Use `"padding_mask"` to indicate values that should not be overridden.
|
153 |
+
"padding_mask": np.array([[1, 1, 1, 1, 0, 0, 0, 0, 0, 0]] * 2),
|
154 |
+
}
|
155 |
+
|
156 |
+
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset(
|
157 |
+
"hf://keras/phi3_mini_128k_instruct_en",
|
158 |
+
preprocessor=None,
|
159 |
+
dtype="bfloat16"
|
160 |
+
)
|
161 |
+
phi3_lm.generate(prompt)
|
162 |
+
```
|
163 |
+
|
164 |
+
Call `fit()` on a single batch.
|
165 |
+
|
166 |
+
```python
|
167 |
+
features = ["The quick brown fox jumped.", "I forgot my homework."]
|
168 |
+
phi3_lm = keras_hub.models.Phi3CausalLM.from_preset("hf://keras/phi3_mini_128k_instruct_en")
|
169 |
+
phi3_lm.fit(x=features, batch_size=2)
|
170 |
+
```
|