-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun.py
53 lines (41 loc) · 1.37 KB
/
run.py
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
#%%
from completions import *
from expand_llm import *
from expand import *
# %%
model, tokenizer, device = load_model()
#%%
# input_text = "The quick brown fox jumpz over"
# input_text = "He asked me to prostate myself before the king"
input_text = "Здравствуйте, я хочу предвыполнить заказ"
inputs: BatchEncoding = tokenize(input_text, tokenizer, device)
#%%
token_probs: list[tuple[int, float]] = calculate_log_probabilities(model, tokenizer, inputs)
#%%
words = split_into_words(token_probs, tokenizer)
log_prob_threshold = -5.0
low_prob_words = [(i, word) for i, word in enumerate(words) if word.logprob < log_prob_threshold]
#%%
contexts = [word.context for _, word in low_prob_words]
#%%
expander = LLMBatchExpander(model, tokenizer)
#%%
series = []
for i, x in enumerate(contexts):
series.append(Series(id=i, tokens=x, budget=5.0))
#%%
batch = Batch(items=series)
#%%
stopping_criterion = create_stopping_criterion_llm(tokenizer)
#%%
expanded = expand(batch, expander, stopping_criterion)
# %%
def print_expansions(expansions: CompletedBatch):
for result in expansions.items:
for expansion in result.expansions:
# convert tokens to string
tokens = [e.token for e in expansion]
s = tokenizer.decode(tokens)
print(f"{result.series.id}: {expansion} {s}")
print_expansions(expanded)
# %%