Commit
•
07029f8
1
Parent(s):
cd2865e
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,63 @@
|
|
1 |
---
|
2 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: mit
|
3 |
+
datasets:
|
4 |
+
- DarwinAnim8or/greentext
|
5 |
+
language:
|
6 |
+
- en
|
7 |
+
tags:
|
8 |
+
- fun
|
9 |
+
- greentext
|
10 |
+
widget:
|
11 |
+
- text: ">be me"
|
12 |
+
example_title: "be me"
|
13 |
+
- text: ">be zoo keeper"
|
14 |
+
co2_eq_emissions:
|
15 |
+
emissions: 30
|
16 |
+
source: "https://mlco2.github.io/impact/#compute"
|
17 |
+
training_type: "fine-tuning"
|
18 |
+
geographical_location: "Oregon, USA"
|
19 |
+
hardware_used: "1x T4, Google Colab"
|
20 |
---
|
21 |
+
|
22 |
+
# GPT-Greentext-1.5b
|
23 |
+
A finetuned version of [GPT2-XL](https://huggingface.co/gpt2-xl) on the 'greentext' dataset.
|
24 |
+
A demo is available [here](https://huggingface.co/spaces/DarwinAnim8or/GPT-Greentext-Playground)
|
25 |
+
The demo playground is recommended over the inference box on the right.
|
26 |
+
|
27 |
+
This is the largest release of the "GPT-Greentext" model series. The other models can be found here:
|
28 |
+
* [355m size model](https://huggingface.co/DarwinAnim8or/GPT-Greentext-355m)
|
29 |
+
* [125m size model](https://huggingface.co/DarwinAnim8or/GPT-Greentext-125m)
|
30 |
+
|
31 |
+
# Training Procedure
|
32 |
+
This was trained on the 'greentext' dataset, on Google Colab.
|
33 |
+
This model was trained for 1 epoch with learning rate 1e-2.
|
34 |
+
Notably this uses the "prompt" and "completion" style jsonl file, rather than the plain text file found in the greentext dataset.
|
35 |
+
This nets somewhat better, mostly more consistent results.
|
36 |
+
|
37 |
+
# Biases & Limitations
|
38 |
+
This likely contains the same biases and limitations as the original GPT2 that it is based on, and additionally heavy biases from the greentext dataset.
|
39 |
+
It should be noted that offensive or not PG-output is definitely possible and likely will happen.
|
40 |
+
|
41 |
+
# Intended Use
|
42 |
+
This model is meant for fun, nothing else.
|
43 |
+
|
44 |
+
# Noteworthy differences between this model and the others
|
45 |
+
This model tends to like no_repeat_ngram_size values of 1 or 2; whereas the other models in this series tend to prefer 3.
|
46 |
+
|
47 |
+
# Sample Use
|
48 |
+
```python
|
49 |
+
#Import model:
|
50 |
+
from happytransformer import HappyGeneration
|
51 |
+
happy_gen = HappyGeneration("GPT2", "DarwinAnim8or/GPT-Greentext-1.5b")
|
52 |
+
|
53 |
+
#Set generation settings:
|
54 |
+
from happytransformer import GENSettings
|
55 |
+
args_top_k = GENSettingsGENSettings(no_repeat_ngram_size=1, do_sample=True, top_k=80, temperature=0.8, max_length=150, early_stopping=False)
|
56 |
+
|
57 |
+
#Generate a response:
|
58 |
+
result = happy_gen.generate_text(""">be me
|
59 |
+
>""", args=args_top_k)
|
60 |
+
|
61 |
+
print(result)
|
62 |
+
print(result.text)
|
63 |
+
```
|