sae_eval/loading_utils.py at main · canrager/sae_eval · GitHub

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
134
135
136
137
138
139
140
141
142
143
144
145
146
147
"""
Partially code from the feature_circuits repo
"""


import os
import re
import json
import random
import torch as t
import torch.nn.functional as F
from dictionary_learning.dictionary import AutoEncoder
from dataclasses import dataclass

@dataclass
class DictionaryCfg():
    def __init__(
        self,
        dictionary_dir,
        dictionary_size
        ) -> None:
        self.dir = dictionary_dir
        self.size = dictionary_size


def load_examples(dataset, num_examples, model, seed=12, pad_to_length=None, length=None):
    examples = []
    dataset_items = open(dataset).readlines()
    random.seed(seed)
    random.shuffle(dataset_items)
    for line in dataset_items:
        data = json.loads(line)
        clean_prefix = model.tokenizer(data["clean_prefix"], return_tensors="pt",
                                        padding=False).input_ids
        patch_prefix = model.tokenizer(data["patch_prefix"], return_tensors="pt",
                                        padding=False).input_ids
        clean_answer = model.tokenizer(data["clean_answer"], return_tensors="pt",
                                        padding=False).input_ids
        patch_answer = model.tokenizer(data["patch_answer"], return_tensors="pt",
                                        padding=False).input_ids
        # only keep examples where answers are single tokens
        if clean_prefix.shape[1] != patch_prefix.shape[1]:
            continue
        # only keep examples where clean and patch inputs are the same length
        if clean_answer.shape[1] != 1 or patch_answer.shape[1] != 1:
            continue
        # if we specify a `length`, filter examples if they don't match
        if length and clean_prefix.shape[1] != length:
            continue
        # if we specify `pad_to_length`, left-pad all inputs to a max length
        prefix_length_wo_pad = clean_prefix.shape[1]
        if pad_to_length:
            model.tokenizer.padding_side = 'right'
            pad_length = pad_to_length - prefix_length_wo_pad
            if pad_length < 0:  # example too long
                continue
            # left padding: reverse, right-pad, reverse
            clean_prefix = t.flip(F.pad(t.flip(clean_prefix, (1,)), (0, pad_length), value=model.tokenizer.pad_token_id), (1,))
            patch_prefix = t.flip(F.pad(t.flip(patch_prefix, (1,)), (0, pad_length), value=model.tokenizer.pad_token_id), (1,))

        example_dict = {"clean_prefix": clean_prefix,
                        "patch_prefix": patch_prefix,
                        "clean_answer": clean_answer.item(),
                        "patch_answer": patch_answer.item(),
                        "annotations": get_annotation(dataset, model, data),
                        "prefix_length_wo_pad": prefix_length_wo_pad,}
        examples.append(example_dict)
        if len(examples) >= num_examples:
            break

    return examples


def load_examples_nopair(dataset, num_examples, model, length=None):
    examples = []
    if isinstance(dataset, str):        # is a path to a .json file
        dataset = json.load(open(dataset))
    elif isinstance(dataset, dict):     # is an already-loaded dictionary
        pass
    else:
        raise ValueError(f"`dataset` is unrecognized type: {type(dataset)}. Must be path (str) or dict")

    max_len = 0     # for padding
    for context_id in dataset:
        context = dataset[context_id]["context"]
        if length is not None and len(context) > length:
            context = context[-length:]
        clean_prefix = model.tokenizer("".join(context), return_tensors="pt",
                        padding=False).input_ids
        max_len = max(max_len, clean_prefix.shape[-1])

    for context_id in dataset:
        answer = dataset[context_id]["answer"]
        context = dataset[context_id]["context"]
        clean_prefix = model.tokenizer("".join(context), return_tensors="pt",
                                    padding=False).input_ids
        clean_answer = model.tokenizer(answer, return_tensors="pt",
                                    padding=False).input_ids
        if clean_answer.shape[1] != 1:
            continue
        prefix_length_wo_pad = clean_prefix.shape[1]
        pad_length = max_len - prefix_length_wo_pad
        # left padding: reverse, right-pad, reverse
        clean_prefix = t.flip(F.pad(t.flip(clean_prefix, (1,)), (0, pad_length), value=model.tokenizer.pad_token_id), (1,))

        example_dict = {"clean_prefix": clean_prefix,
                        "clean_answer": clean_answer.item(),
                        "prefix_length_wo_pad": prefix_length_wo_pad,}
        examples.append(example_dict)
        if len(examples) >= num_examples:
            break

    return examples

def get_annotation(dataset, model, data):
    # First, understand which dataset we're working with
    structure = None
    if "within_rc" in dataset:
        structure = "within_rc"
        template = "the_subj subj_main that the_dist subj_dist"
    elif "rc.json" in dataset or "rc_" in dataset:
        structure = "rc"
        template = "the_subj subj_main that the_dist subj_dist verb_dist"
    elif "simple.json" in dataset or "simple_" in dataset:
        structure = "simple"
        template = "the_subj subj_main"
    elif "nounpp.json" in dataset or "nounpp_" in dataset:
        structure = "nounpp"
        template = "the_subj subj_main prep the_dist subj_dist"

    if structure is None:
        return {}

    annotations = {}

    # Iterate through words in the template and input. Get token spans
    curr_token = 0
    for template_word, word in zip(template.split(), data["clean_prefix"].split()):
        if word != "The":
            word = " " + word
        word_tok = model.tokenizer(word, return_tensors="pt", padding=False).input_ids
        num_tokens = word_tok.shape[1]
        span = (curr_token, curr_token + num_tokens-1)
        curr_token += num_tokens
        annotations[template_word] = span

    return annotations