Spaces:
Running
Running
File size: 6,466 Bytes
25f66ac 0cfd419 25f66ac 50d13b1 25f66ac 50d13b1 16902fa 50d13b1 16902fa 50d13b1 c671ff0 50d13b1 25f66ac 50d13b1 25f66ac c24faa7 25f66ac ceb140a c24faa7 25f66ac c24faa7 25f66ac 50d13b1 25f66ac 50d13b1 c24faa7 25f66ac 50d13b1 25f66ac |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 |
import gradio as gr
import datetime
import json
import requests
from constants import *
def process(query_type, index_desc, **kwargs):
timestamp = datetime.datetime.now().strftime('%Y%m%d-%H%M%S')
index = INDEX_BY_DESC[index_desc]
data = {
'source': 'hf' if not DEBUG else 'hf-dev',
'timestamp': timestamp,
'query_type': query_type,
'index': index,
}
data.update(kwargs)
print(json.dumps(data))
if API_URL is None:
raise ValueError(f'API_URL envvar is not set!')
try:
response = requests.post(API_URL, json=data, timeout=MAX_TIMEOUT_IN_SECONDS)
except requests.exceptions.Timeout:
raise ValueError('Web request timed out. Please try again later.')
except requests.exceptions.RequestException as e:
raise ValueError(f'Web request error: {e}')
if response.status_code == 200:
result = response.json()
else:
raise ValueError(f'HTTP error {response.status_code}: {response.json()}')
if DEBUG:
print(result)
return result
def creativity(index_desc, query):
result = process('creativity', index_desc, query=query)
latency = '' if 'latency' not in result else f'{result["latency"]:.3f}'
if 'error' in result:
ci = result['error']
htmls = [''] * (NGRAM_LEN_MAX - NGRAM_LEN_MIN + 1)
return tuple([latency, ci] + htmls)
rs = result['rs']
tokens = result['tokens']
highlighteds_by_n = {}
uniqueness_by_n = {}
for n in range(NGRAM_LEN_MIN, NGRAM_LEN_MAX + 1):
highlighteds = [False] * len(tokens)
last_r = 0
for l, r in enumerate(rs):
if r - l < n:
continue
for i in range(max(last_r, l), r):
highlighteds[i] = True
last_r = r
uniqueness = sum([1 for h in highlighteds if not h]) / len(highlighteds)
highlighteds_by_n[n] = highlighteds
uniqueness_by_n[n] = uniqueness
ci = sum(uniqueness_by_n.values()) / len(uniqueness_by_n)
ci = f'{ci:.2%}'
htmls = []
for n in range(NGRAM_LEN_MIN, NGRAM_LEN_MAX + 1):
html = ''
highlighteds = highlighteds_by_n[n]
line_len = 0
for i, (token, highlighted) in enumerate(zip(tokens, highlighteds)):
if line_len >= MAX_DISP_CHARS_PER_LINE and token.startswith('β'):
html += '<br/>'
line_len = 0
color = '(255, 128, 128, 0.5)'
if token == '<0x0A>':
disp_token = '\\n'
is_linebreak = True
else:
disp_token = token.replace('β', ' ')
is_linebreak = False
if highlighted:
html += f'<span id="hldoc-token-{i}" style="background-color: rgba{color};" class="background-color: rgba{color};">{disp_token}</span>'
else:
html += disp_token
if is_linebreak:
html += '<br/>'
line_len = 0
else:
line_len += len(token)
html = '<div><p id="hldoc" style="font-size: 16px;">' + html.strip(' ') + '</p></div>'
htmls.append(html)
return tuple([latency, ci] + htmls)
with gr.Blocks() as demo:
with gr.Column():
gr.HTML(
f'''<h1 text-align="center">Creativity Index</h1>
<p style='font-size: 16px;'>Compute the <a href="https://arxiv.org/pdf/2410.04265">Creativity Index</a> of a piece of text.</p>
<p style='font-size: 16px;'>The Creativity Index is computed based on verbatim matching against massive text corpora and is powered by <a href="https://infini-gram.io">infini-gram</a>. It is defined as the ratio of tokens not covered by n-grams (n >= L) that can be found in the corpus, averaged across {NGRAM_LEN_MIN} <= L <= {NGRAM_LEN_MAX}. You can view the covered tokens (highlighted in red background) for each value of L.</p>
<p style='font-size: 16px;'><b>Note:</b> The input text is limited to {MAX_QUERY_CHARS} characters. Each query has a timeout of {MAX_TIMEOUT_IN_SECONDS} seconds. If you have waited 30 seconds and receive an error, you can try submitted the same query and it's more likely to work on the second try.</p>
<p style='font-size: 16px;'><b>Disclaimer 1:</b> The Creativity Index of text that appear exactly in the corpora may be deflated. In our paper, we remove exact duplicates (including quotations and citations) from the corpus before computing the Creativity Index. However, deduplication is not applied in this demo.</p>
<p style='font-size: 16px;'><b>Disclaimer 2:</b> The Creativity Index of text generated by latest models (e.g., GPT-4) may be inflated. This is because we don't have all the data that these models are trained on, and our supported corpora have a earlier cutoff date (Dolma-v1.7 is Oct 2023, RedPajama is Mar 2023, Pile is 2020).</p>
'''
)
with gr.Row():
with gr.Column(scale=1, min_width=240):
index_desc = gr.Radio(choices=INDEX_DESCS, label='Corpus', value=INDEX_DESCS[2])
with gr.Column(scale=3):
creativity_query = gr.Textbox(placeholder='Enter a piece of text here', label='Input', interactive=True, lines=10)
with gr.Row():
creativity_clear = gr.ClearButton(value='Clear', variant='secondary', visible=True)
creativity_submit = gr.Button(value='Submit', variant='primary', visible=True)
creativity_latency = gr.Textbox(label='Latency (milliseconds)', interactive=False, lines=1)
with gr.Column(scale=4):
creativity_ci = gr.Label(value='', label='Creativity Index')
creativity_htmls = []
for n in range(NGRAM_LEN_MIN, NGRAM_LEN_MAX + 1):
with gr.Tab(label=f'L={n}'):
creativity_htmls.append(gr.HTML(value='', label=f'L={n}'))
creativity_clear.add([creativity_query, creativity_latency, creativity_ci] + creativity_htmls)
creativity_submit.click(creativity, inputs=[index_desc, creativity_query], outputs=[creativity_latency, creativity_ci] + creativity_htmls, api_name=False)
demo.queue(
default_concurrency_limit=DEFAULT_CONCURRENCY_LIMIT,
max_size=MAX_SIZE,
api_open=False,
).launch(
max_threads=MAX_THREADS,
debug=DEBUG,
show_api=False,
)
|