aapot commited on
Commit
58ed682
1 Parent(s): f71d1c2

Add toxicity calculation script

Browse files
Files changed (1) hide show
  1. calculate_toxicity_labels.py +61 -0
calculate_toxicity_labels.py ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer, FlaxBertForSequenceClassification
2
+ import datasets
3
+ import jax
4
+ import jax.numpy as jnp
5
+ import time
6
+ from flax.training.common_utils import shard
7
+ from jax import pmap
8
+
9
+
10
+ def pred_fn(inputs):
11
+ outputs = model(**inputs)
12
+ return jax.nn.sigmoid(outputs.logits)
13
+
14
+
15
+ def get_toxicity(batch, batch_size):
16
+ num_examples = len(batch["text"])
17
+ inputs = tokenizer(
18
+ batch["text"],
19
+ return_tensors="np",
20
+ truncation=True,
21
+ padding="max_length",
22
+ max_length=512,
23
+ )
24
+
25
+ inputs = shard(
26
+ {
27
+ k: jnp.pad(jnp.array(v), ((0, batch_size - num_examples), (0, 0)))
28
+ for k, v in inputs.items()
29
+ }
30
+ )
31
+ preds = p_pred(inputs)
32
+ preds = preds.reshape(-1, preds.shape[-1])[:num_examples]
33
+ for k, v in model.config.id2label.items():
34
+ batch[v] = preds[:, k].tolist()
35
+ return batch
36
+
37
+
38
+ p_pred = pmap(pred_fn, "inputs")
39
+
40
+ tokenizer = AutoTokenizer.from_pretrained("TurkuNLP/bert-large-finnish-cased-toxicity")
41
+ model = FlaxBertForSequenceClassification.from_pretrained(
42
+ "TurkuNLP/bert-large-finnish-cased-toxicity", from_pt=True, dtype=jnp.bfloat16
43
+ )
44
+
45
+
46
+ dataset = datasets.load_from_disk("/researchdisk/mc4_3.1.0_fi_cleaned")
47
+
48
+ BATCH_SIZE = 8192
49
+ dataset = dataset.map(
50
+ get_toxicity,
51
+ num_proc=1,
52
+ batched=True,
53
+ batch_size=BATCH_SIZE,
54
+ fn_kwargs={"batch_size": BATCH_SIZE},
55
+ )
56
+ print(dataset)
57
+
58
+ # SAVE DATASET
59
+ dataset.save_to_disk(
60
+ "/researchdisk/mc4_3.1.0_fi_cleaned_dataset_toxicity_labels", num_proc=32
61
+ )