anonymous-aardvark
commited on
Commit
•
4aa3dc9
1
Parent(s):
85ecf02
welcome formatting
Browse files- demo_watermark.py +5 -13
demo_watermark.py
CHANGED
@@ -482,19 +482,11 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
|
|
482 |
"""
|
483 |
This demo uses open-source language models. Today, these models are less powerful than proprietary commercial tools like ChatGPT, Claude, Bard, or Bing/Sydney.
|
484 |
|
485 |
-
Models like [BLOOM (175B)](https://huggingface.co/bigscience/bloom) are designed to "complete" your prompt, and are not fine-tuned to follow instructions.
|
486 |
-
|
487 |
-
Some
|
488 |
-
|
489 |
-
|
490 |
-
Some of the models available in this demo are fine-tuned to follow instructions but have different strengths and will showcase different
|
491 |
-
types of watermark behavior. [BLOOMZ](https://huggingface.co/bigscience/bloomz) is an instruction tuned variant of BLOOM capable of following instructions in dozens of languages zero-shot
|
492 |
-
and can generate long and coherent paragraphs and stories given the right prompt.
|
493 |
-
The FLAN models [FLAN-t5-xxl (11B)](https://huggingface.co/google/flan-t5-xxl) and [FLAN-UL2 (20B)](https://huggingface.co/google/flan-ul2) are fine-tuned on a variety of in-context few-shot learning NLP tasks,
|
494 |
-
such as reasoning, and question answering.
|
495 |
-
|
496 |
-
Generally, short, low entropy scenarios where the model has very few choices in terms of correct/suitable responses to the prompt
|
497 |
-
will not exhibit as strong of a watermark presence, while longer watermarked outputs will produce higher detection statistics.
|
498 |
"""
|
499 |
)
|
500 |
gr.Markdown(
|
|
|
482 |
"""
|
483 |
This demo uses open-source language models. Today, these models are less powerful than proprietary commercial tools like ChatGPT, Claude, Bard, or Bing/Sydney.
|
484 |
|
485 |
+
Models like [BLOOM (175B)](https://huggingface.co/bigscience/bloom) are designed to "complete" your prompt, and are not fine-tuned to follow instructions. For best results, prompt that model with a few sentences that form the beginning of a paragraph, and then allow it to "continue" your paragraph. Some examples include the opening paragraph of a wikipedia article, or the first few sentences of a story. Longer prompts that end mid-sentence will result in more fluent generations.
|
486 |
+
|
487 |
+
Some of the models available in this demo are fine-tuned to follow instructions but have different strengths and will showcase differenttypes of watermark behavior. [BLOOMZ](https://huggingface.co/bigscience/bloomz) is an instruction tuned variant of BLOOM capable of following instructions in dozens of languages zero-shot and can generate long and coherent paragraphs and stories given the right prompt. The FLAN models [FLAN-t5-xxl (11B)](https://huggingface.co/google/flan-t5-xxl) and [FLAN-UL2 (20B)](https://huggingface.co/google/flan-ul2) are fine-tuned on a variety of in-context few-shot learning NLP tasks, such as reasoning, and question answering.
|
488 |
+
|
489 |
+
Generally, short, low entropy scenarios where the model has very few choices in terms of correct/suitable responses to the prompt will not exhibit as strong of a watermark presence, while longer watermarked outputs will produce higher detection statistics.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
490 |
"""
|
491 |
)
|
492 |
gr.Markdown(
|