SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2010.05607 · 2020

The elephant in the interpretability room: Why use attention as explanation when we have saliency methods?

arXiv · PDF · Open in the Atlas

Code that ran

11 of 16 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 2 have no contract verdict; 8 failures of our instrument; 5 no recorded run); 1 of them was also run on the host, not in the jail, by the peer-agreement lane that set its level, measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2010.05607.json · badge: /api/ran/2010.05607.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
jessevig/bertviz canonical (classifier label)
classifier basis: census
11 ran · 5 no recorded run
FunctionStatusWhere it lives
swish Ran
ran; our contract draft was wrong, not the code
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_bert.py
code served (permissive licence) · get_code("0f786c407fb1ee4c")
How it ran

Ran in Syntology's jail on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. The comparison behind this sample's level is a different run: the peer-agreement lane ran it again on the host, not in the jail (report dated 2026-09-04T22:33:52Z). If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# The function swish takes a single parameter 'x' which is used with:
# - torch.sigmoid(x) - requires a tensor
# - x * ... - requires a tensor that supports multiplication
# So x should be a tensor

x = torch.randn(2, 4)

ARGS = {
    "x": x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
build_tf_to_pytorch_map Ran
ran; our contract draft was wrong, not the code
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_transfo_xl.py
code served (permissive licence) · get_code("35ef111fc00c21b4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class Config:
    def __init__(self):
        self.tie_projs = [False]
        self.tie_weight = True
        self.untie_r = False

class LayerNorm(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.weight = nn.Parameter(torch.randn(dim))
        self.bias = nn.Parameter(torch.randn(dim))

class DecAttn(nn.Module):
    def __init__(self, d_model=4, d_head=2):
        super().__init__()
        self.layer_norm = LayerNorm(d_model)
        self.o_net = nn.Linear(d_model, d_model)
        self.qkv_net = nn.Linear(d_model, 3 * d_model)
        self.r_net = nn.Linear(d_model, d_model)
        self.r_r_bias = nn.Parameter(torch.randn(d_model))
        self.r_w_bias = nn.Parameter(torch.randn(d_model))

class PosFeedForward(nn.Module):
    def __init__(self, d_model=4, d_ff=8):
        super().__init__()
        self.layer_norm = LayerNorm(d_model)
        self.CoreNet = nn.Sequential(
            nn.Linear(d_model, d_ff),
            nn.ReLU(),
            nn.Linear(d_ff, d_ff),
            nn.Linear(d_ff, d_model)
        )

class TransformerLayer(nn.Module):
    def __init__(self, d_model=4):
        super().__init__()
        self.dec_attn = DecAttn(d_model)
        self.pos_ff = PosFeedForward(d_model)

class WordEmb(nn.Module):
    def __init__(self, vocab_size=100, d_model=4, n_cutoffs=1):
        super().__init__()
        self.emb_layers = nn.ModuleList([nn.Embedding(vocab_size, d_model) for _ in range(n_cutoffs)])
        self.emb_projs = nn.ModuleList([nn.Linear(d_model, d_model) for _ in range(n_cutoffs)])

class AdaptiveSoftmax(nn.Module):
    def __init__(self, vocab_size=100, d_model=4, n_cutoffs=1):
        super().__init__()
        self.cluster_weight = nn.Parameter(torch.randn(n_cutoffs, d_model))
        self.cluster_bias = nn.Parameter(torch.randn(n_cutoffs))
        self.out_layers = nn.ModuleList([nn.Linear(d_model, vocab_size) for _ in range(n_cutoffs)])
        self.out_projs = nn.ModuleList([nn.Linear(d_model, d_model) for _ in range(n_cutoffs)])

class Transformer(nn.Module):
    def __init__(self, d_model=4, n_layers=2, n_cutoffs=1):
        super().__init__()
        self.word_emb = WordEmb(vocab_size=100, d_model=d_model, n_cutoffs=n_cutoffs)
        self.layers = nn.ModuleList([TransformerLayer(d_model) for _ in range(n_layers)])
        self.r_r_bias = nn.Parameter(torch.randn(d_model))
        self.r_w_bias = nn.Parameter(torch.randn(d_model))

class Model(nn.Module):
    def __init__(self, d_model=4, n_layers=2, n_cutoffs=1):
        super().__init__()
        self.transformer = Transformer(d_model=d_model, n_layers=n_layers, n_cutoffs=n_cutoffs)
        self.crit = AdaptiveSoftmax(vocab_size=100, d_model=d_model, n_cutoffs=n_cutoffs)

config = Config()
model = Model(d_model=4, n_layers=2, n_cutoffs=1)

ARGS = {
    'model': model,
    'config': config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
build_tf_xlnet_to_pytorch_map Ran
ran on a synthesized input
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_xlnet.py
code served (permissive licence) · get_code("c92ba449e48ecffe")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Build minimal attention layer
class RelAttn(nn.Module):
    def __init__(self, d_model=4, num_heads=2):
        super().__init__()
        self.layer_norm = nn.LayerNorm(d_model)
        self.o = nn.Linear(d_model, d_model)
        self.q = nn.Linear(d_model, d_model)
        self.k = nn.Linear(d_model, d_model)
        self.r = nn.Linear(d_model, d_model)
        self.v = nn.Linear(d_model, d_model)
        self.r_r_bias = torch.randn(4)
        self.r_w_bias = torch.randn(4)
        self.r_s_bias = torch.randn(4)
        self.seg_embed = torch.randn(2, 4)

# Build minimal feedforward layer
class FFN(nn.Module):
    def __init__(self, d_model=4, d_ff=8):
        super().__init__()
        self.layer_norm = nn.LayerNorm(d_model)
        self.layer_1 = nn.Linear(d_model, d_ff)
        self.layer_2 = nn.Linear(d_ff, d_model)

# Build minimal transformer layer
class TransformerLayer(nn.Module):
    def __init__(self, d_model=4, d_ff=8):
        super().__init__()
        self.rel_attn = RelAttn(d_model=d_model)
        self.ff = FFN(d_model=d_model, d_ff=d_ff)

# Build minimal transformer
class Transformer(nn.Module):
    def __init__(self, num_layers=2, d_model=4, d_ff=8, untie_r=False):
        super().__init__()
        self.word_embedding = nn.Embedding(100, d_model)
        self.mask_emb = torch.randn(d_model)
        self.layer = nn.ModuleList([TransformerLayer(d_model, d_ff) for _ in range(num_layers)])
        self.untie_r = untie_r
        
        if not untie_r:
            self.r_r_bias = torch.randn(d_model)
            self.r_w_bias = torch.randn(d_model)
            self.r_s_bias = torch.randn(d_model)
            self.seg_embed = torch.randn(2, d_model)

# Build full model with optional lm_loss, sequence_summary, logits_proj
class XLNetModel(nn.Module):
    def __init__(self, num_layers=2, d_model=4, d_ff=8, untie_r=False, with_lm_loss=True, with_sequence_summary=False, with_logits_proj=False):
        super().__init__()
        self.transformer = Transformer(num_layers=num_layers, d_model=d_model, d_ff=d_ff, untie_r=untie_r)
        if with_lm_loss:
            self.lm_loss = nn.Linear(d_model, 100)
        if with_sequence_summary:
            self.sequence_summary = nn.Module()
            self.sequence_summary.summary = nn.Linear(d_model, d_model)
        if with_logits_proj:
            self.logits_proj = nn.Linear(d_model, 1)

# Build config
class Config:
    def __init__(self, untie_r=False, finetuning_task=None):
        self.untie_r = untie_r
        self.finetuning_task = finetuning_task

# Create instances
model = XLNetModel(num_layers=2, d_model=4, d_ff=8, untie_r=False, with_lm_loss=True, with_sequence_summary=True, with_logits_proj=True)
config = Config(untie_r=False, finetuning_task="regression")
tf_weights = {
    'model/sequnece_summary/summary/kernel': True,
    'model/regression_regression/logit/kernel': True,
}

ARGS = {
    "model": model,
    "config": config,
    "tf_weights": tf_weights,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
format_special_chars Ran
ran; our contract draft was wrong, not the code
jessevig/bertviz/bertviz/neuron_view.py
code served (permissive licence) · get_code("11442089a6433797")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# The function format_special_chars takes a list of tokens (strings)
# and replaces special characters with spaces
# tokens parameter: a list of strings

tokens = [
    'Ġhello',
    'world▁test',
    'Ġ▁mixed',
    'normal'
]

ARGS = {
    'tokens': tokens
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
gelu Ran
ran, honoured the contract we drafted
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_bert.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from tanmay1618/text_to_sql_bert, is pointer only (licence: NONE) · get_code("fdc64f4c72036ae4")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from tanmay1618/text_to_sql_bert, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_attention Ran
ran; our contract draft was wrong, not the code
jessevig/bertviz/bertviz/neuron_view.py
code served (permissive licence) · get_code("d4eae117e16a3f41")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from collections import defaultdict

torch.manual_seed(0)

# Create a minimal tokenizer stand-in
class MinimalTokenizer:
    def __init__(self):
        self.cls_token = '[CLS]'
        self.sep_token = '[SEP]'
        self.vocab_size = 100
    
    def tokenize(self, text):
        # Simple whitespace tokenization
        return text.split()
    
    def convert_tokens_to_ids(self, tokens):
        # Return dummy token IDs
        return [hash(t) % self.vocab_size for t in tokens]

# Create a minimal BERT-like model that outputs attention
class MinimalBertModel(nn.Module):
    def __init__(self, num_layers=2, num_heads=4, hidden_size=8):
        super().__init__()
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.hidden_size = hidden_size
        self.embedding = nn.Embedding(100, hidden_size)
        self.layers = nn.ModuleList([
            nn.Linear(hidden_size, hidden_size) for _ in range(num_layers)
        ])
    
    def forward(self, input_ids, token_type_ids=None):
        batch_size, seq_len = input_ids.shape
        
        # Embed input
        x = self.embedding(input_ids)
        
        # Generate attention outputs for each layer
        attn_outputs = []
        for layer_idx in range(self.num_layers):
            # Create dummy attention: [batch_size, num_heads, seq_len, seq_len]
            attn = torch.randn(batch_size, self.num_heads, seq_len, seq_len)
            attn = torch.softmax(attn, dim=-1)
            
            # Create dummy queries and keys for include_queries_and_keys
            queries = torch.randn(batch_size, self.num_heads, seq_len, self.hidden_size // self.num_heads)
            keys = torch.randn(batch_size, self.num_heads, seq_len, self.hidden_size // self.num_heads)
            
            attn_outputs.append({
                'attn': attn,
                'queries': queries,
                'keys': keys,
            })
        
        # Return tuple where last element is attention list
        return (x, attn_outputs)

# Construct the arguments
tokenizer = MinimalTokenizer()
model = MinimalBertModel(num_layers=2, num_heads=3)
model.eval()

ARGS = {
    'model': model,
    'model_type': 'bert',
    'tokenizer': tokenizer,
    'sentence_a': 'hello world test',
    'sentence_b': None,
    'include_queries_and_keys': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_masks Ran
ran on a synthesized input
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_xlm.py
code served (permissive licence) · get_code("f5d8d44f65a08b76")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 3, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# slen: sequence length (int)
# lengths: tensor of actual sequence lengths for each batch item
# causal: boolean for whether to use causal masking
# padding_mask: optional pre-computed padding mask

slen = 5
bs = 2

# lengths should have shape (bs,) with values <= slen
lengths = torch.tensor([3, 4], dtype=torch.long)

# causal mask flag
causal = False

# padding_mask is optional, set to None to let the function compute it
padding_mask = None

ARGS = {
    "slen": slen,
    "lengths": lengths,
    "causal": causal,
    "padding_mask": padding_mask,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 3, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_tf_weights_in_bert Ran
ran; our contract draft was wrong, not the code
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_bert.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from tanmay1618/text_to_sql_bert, is pointer only (licence: NONE) · get_code("0c145728a35789fa")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from tanmay1618/text_to_sql_bert, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

prune_conv1d_layer Ran
ran; our contract draft was wrong, not the code
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_utils.py
code served (permissive licence) · get_code("7183a17dd79327a8")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Conv1D. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

class Conv1D(nn.Module):
    def __init__(self, nf, nx):
        super(Conv1D, self).__init__()
        self.nf = nf
        w = torch.empty(nx, nf)
        nn.init.normal_(w, std=0.02)
        self.weight = nn.Parameter(w)
        self.bias = nn.Parameter(torch.zeros(nf))

    def forward(self, x):
        size_out = x.size()[:-1] + (self.nf,)
        x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
        x = x.view(*size_out)
        return x

# Create a Conv1D layer to prune
# nf=5 (output features), nx=4 (input features)
layer = Conv1D(nf=5, nx=4)

# Create an index tensor to select which dimensions to keep
# dim=1 means we're pruning the output dimension
# Select indices [0, 2] from the output features
index = torch.tensor([0, 2], dtype=torch.long)

# dim parameter: prune along dimension 1 (output features)
dim = 1

ARGS = {
    "layer": layer,
    "index": index,
    "dim": dim
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Conv1D — a Conv1D output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Conv1D output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
prune_linear_layer Ran
ran; our contract draft was wrong, not the code
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_utils.py
code served (permissive licence) · get_code("d30d6c3098df2c42")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1) and returned Linear. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from torch import nn

torch.manual_seed(0)

# Create a linear layer to prune
# Using distinct sizes: input_features=3, output_features=4
layer = nn.Linear(in_features=3, out_features=4, bias=True)

# Create an index tensor to select which output features to keep
# Selecting 2 out of 4 output features
index = torch.tensor([0, 2], dtype=torch.long)

# dim=0 means we're pruning along the output dimension (rows of weight matrix)
dim = 0

ARGS = {
    "layer": layer,
    "index": index,
    "dim": dim,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Linear — a Linear output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Linear output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sample_logits Ran
ran; our contract draft was wrong, not the code
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_transfo_xl_utilities.py
code served (permissive licence) · get_code("93228a3eb5c4d179")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 3, dtype family float, shape (2, 3, 6) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Small embedding layer
n_vocab = 20
n_emb = 4
embedding = nn.Embedding(n_vocab, n_emb)

# Bias vector
bias = torch.randn(n_vocab)

# Labels: [b1, b2] with valid vocab indices
b1, b2 = 2, 3
labels = torch.randint(0, n_vocab, (b1, b2))

# Inputs: [b1, b2, n_emb]
inputs = torch.randn(b1, b2, n_emb)

# Sampler that implements the sample() method
class LogUniformSampler:
    def __init__(self, n_vocab, n_sample=5):
        self.n_vocab = n_vocab
        self.n_sample = n_sample
    
    def sample(self, labels):
        # Sample negative samples
        neg_samples = torch.randint(0, self.n_vocab, (self.n_sample,))
        
        # Log probabilities
        true_log_probs = torch.randn(labels.size(0), labels.size(1))
        samp_log_probs = torch.randn(self.n_sample)
        
        return true_log_probs, samp_log_probs, neg_samples

sampler = LogUniformSampler(n_vocab=n_vocab, n_sample=5)

ARGS = {
    "embedding": embedding,
    "bias": bias,
    "labels": labels,
    "inputs": inputs,
    "sampler": sampler
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 3, dtype family float, shape (2, 3, 6) — shape (2, 3, 6) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_tf_weights_in_gpt2 Not yet run
unverified (harvested, no recorded run)
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_gpt2.py
code served (permissive licence) · get_code("654e91efc679fe84")
How it ran

no invocation recorded.

load_tf_weights_in_openai_gpt Not yet run
unverified (harvested, no recorded run)
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_openai.py
code served (permissive licence) · get_code("42ee168c3ae12e60")
How it ran

no invocation recorded.

load_tf_weights_in_transfo_xl Not yet run
unverified (harvested, no recorded run)
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_transfo_xl.py
code served (permissive licence) · get_code("23d246e59ef9447b")
How it ran

no invocation recorded.

load_tf_weights_in_xlnet Not yet run
unverified (harvested, no recorded run)
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_xlnet.py
code served (permissive licence) · get_code("4a94662ff3b5c5cf")
How it ran

no invocation recorded.

prune_layer Not yet run
unverified (harvested, no recorded run)
jessevig/bertviz/bertviz/transformers_neuron_view/modeling_utils.py
code served (permissive licence) · get_code("e31093ffc8275cff")
How it ran

no invocation recorded.

Also present in other papers' code

Of 16 samples lifted from this paper's linked repositories, 16 carry a function identity and 0 do not. Of the 16 distinct functions with an identity, 16 are also present in other papers whose code we hold (16 with at least one paper dated earlier, 3 the same month, 12 later) and 0 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
swish 206 other papers
47 dated earlier, 5 the same month, 153 later, 1 same year, order not determinable · 44 only as an assembly brick
1408.5882 (2014-08)
Convolutional Neural Networks for Sentence Classification
1805.12114 (2018-05)
Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models
1809.01357 (2018-09)
Zero Shot Learning for Code Education: Rubric Sampling with Deep Learning Inference
and 203 more not named here
gelu 198 other papers
57 dated earlier, 4 the same month, 135 later, 2 same year, order not determinable · 32 only as an assembly brick
1408.5882 (2014-08)
Convolutional Neural Networks for Sentence Classification
1810.04805 (2018-10)
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
1901.10125 (2019-01)
Glyce: Glyph-vectors for Chinese Character Representations
and 195 more not named here
prune_linear_layer 29 other papers
14 dated earlier, 14 later, 1 same year, order not determinable
1810.04805 (2018-10)
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
1906.08230 (2019-06)
Evaluating Protein Transfer Learning with TAPE
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
and 26 more not named here
load_tf_weights_in_bert 19 other papers
8 dated earlier, 1 the same month, 10 later
1810.04805 (2018-10)
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
1902.01069 (2019-02)
A Comprehensive Exploration on WikiSQL with Table-Aware Word Contextualization
1909.11942 (2019-09)
ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
and 16 more not named here
prune_conv1d_layer 16 other papers
8 dated earlier, 8 later
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1909.05311 (2019-09)
Graph-Based Reasoning over Heterogeneous External Knowledge for Commonsense Question Answering
1909.11942 (2019-09)
ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
and 13 more not named here
sample_logits 14 other papers
10 dated earlier, 4 later · 2 only as an assembly brick
1901.02860 (2019-01)
Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1908.09355 (2019-08)
Patient Knowledge Distillation for BERT Model Compression
and 11 more not named here
build_tf_to_pytorch_map 10 other papers
9 dated earlier, 1 later
1901.02860 (2019-01)
Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1908.09355 (2019-08)
Patient Knowledge Distillation for BERT Model Compression
and 7 more not named here
load_tf_weights_in_openai_gpt 8 other papers
5 dated earlier, 3 later
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1909.05311 (2019-09)
Graph-Based Reasoning over Heterogeneous External Knowledge for Commonsense Question Answering
2002.01808 (2020-02)
K-Adapter: Infusing Knowledge into Pre-Trained Models with Adapters
and 5 more not named here
prune_layer 7 other papers
4 dated earlier, 3 later
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1909.11942 (2019-09)
ALBERT: A Lite BERT for Self-supervised Learning of Language Representations
2005.00558 (2020-05)
POINTER: Constrained Progressive Text Generation via Insertion-based Generative Pre-training
and 4 more not named here
load_tf_weights_in_gpt2 6 other papers
2 dated earlier, 4 later
2002.01808 (2020-02)
K-Adapter: Infusing Knowledge into Pre-Trained Models with Adapters
2004.03829 (2020-04)
Exploring Versatile Generative Language Model Via Parameter-Efficient Transfer Learning
2201.06885 (2022-01)
Evidence-aware Fake News Detection with Graph Neural Networks
and 3 more not named here
load_tf_weights_in_transfo_xl 6 other papers
5 dated earlier, 1 later
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1909.05311 (2019-09)
Graph-Based Reasoning over Heterogeneous External Knowledge for Commonsense Question Answering
2002.01808 (2020-02)
K-Adapter: Infusing Knowledge into Pre-Trained Models with Adapters
and 3 more not named here
build_tf_xlnet_to_pytorch_map 4 other papers
3 dated earlier, 1 later
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1909.05311 (2019-09)
Graph-Based Reasoning over Heterogeneous External Knowledge for Commonsense Question Answering
2007.02439 (2020-07)
Pretrained Generalized Autoregressive Model with Adaptive Probabilistic Label Clusters for Extreme Multi-label Text Classification
and 1 more not named here
get_masks 3 other papers
3 dated earlier
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1909.05311 (2019-09)
Graph-Based Reasoning over Heterogeneous External Knowledge for Commonsense Question Answering
2005.00558 (2020-05)
POINTER: Constrained Progressive Text Generation via Insertion-based Generative Pre-training
load_tf_weights_in_xlnet 3 other papers
3 dated earlier
1906.08237 (2019-06)
XLNet: Generalized Autoregressive Pretraining for Language Understanding
1909.05311 (2019-09)
Graph-Based Reasoning over Heterogeneous External Knowledge for Commonsense Question Answering
2007.02439 (2020-07)
Pretrained Generalized Autoregressive Model with Adaptive Probabilistic Label Clusters for Extreme Multi-label Text Classification
format_special_chars 1 other paper
1 dated earlier
1906.05714 (2019-06)
A Multiscale Visualization of Attention in the Transformer Model
get_attention 1 other paper
1 dated earlier
1906.05714 (2019-06)
A Multiscale Visualization of Attention in the Transformer Model

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

There is a recent surge of interest in using attention as explanation of model predictions, with mixed evidence on whether attention can be used as such. While attention conveniently gives us one weight per input token and is easily extracted, it is often unclear toward what goal it is used as explanation. We find that often that goal, whether explicitly stated or not, is to find out what input tokens are the most relevant to a prediction, and that the implied user for the explanation is a model developer. For this goal and user, we argue that input saliency methods are better suited, and that there are no compelling reasons to use attention, despite the coincidence that it provides a weight for each input. With this position paper, we hope to shift some of the recent focus on attention to saliency methods, and for authors to clearly state the goal and user for their explanations.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2010.05607")
get_code_for_paper("2010.05607")
have("2010.05607")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2010.05607.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2010.05607.svg)](https://syntology.ai/paper/2010.05607)

Connect an agent — have() is free.