SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2201.02177 · 2022

Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasets

arXiv · PDF · Open in the Atlas

Code that ran

11 of 14 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 8 have no contract verdict; 2 failures of our instrument; 3 no recorded run), measured 2026-11-21; 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/2201.02177.json · badge: /api/ran/2201.02177.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
openai/grok canonical (classifier label)
classifier basis not recorded on the link
8 ran · 1 no recorded run
ironjr/grokfast extension (classifier label)
classifier basis: classifier_2025
2 ran · 0 no recorded run
danielmamay/grokking community (classifier label)
classifier basis not recorded on the link
1 ran · 2 no recorded run
FunctionStatusWhere it lives
compute_measure Ran
ran on a synthesized input
openai/grok/grok/metrics.py
code served (permissive licence) · get_code("8e9ee805c0204396")
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 float, ndim 0, dtype family float, shape () 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
import math

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = nn.Linear(4, 3)
        self.linear2 = nn.Linear(3, 2)
    
    def forward(self, x):
        x = self.linear1(x)
        x = self.linear2(x)
        return x

def simple_measure_func(module, init_module, **kwargs):
    """
    Simple measure function that computes the L2 norm of weight difference.
    """
    if isinstance(module, nn.Linear):
        weight_diff = (module.weight - init_module.weight).norm()
        return weight_diff.item() + 1e-8  # Add small epsilon to avoid log(0)
    return 1e-8

# Create model and init_model
model = SimpleModel()
init_model = SimpleModel()

# Store initial weights
for m, im in zip(model.modules(), init_model.modules()):
    if isinstance(m, nn.Linear):
        im.weight.data = m.weight.data.clone()
        im.bias.data = m.bias.data.clone()

# Simulate training by modifying weights slightly
for param in model.parameters():
    param.data += torch.randn_like(param) * 0.1

# Define ARGS
ARGS = {
    'model': model,
    'init_model': init_model,
    'measure_func': simple_measure_func,
    'operator': 'sum',
    'kwargs': {},
    'p': 2
}
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 float, ndim 0, dtype family float, shape () — shape () 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
factor_expts Ran
ran on a synthesized input
openai/grok/grok/visualization.py
code served (permissive licence) · get_code("e95fa4ee731c03ea")
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

torch.manual_seed(0)

# The function factor_expts takes a list of experiment name strings
# It parses each string by splitting on '_', extracts architecture and time parameters
# The format appears to be: word_word_word_word-t_...
# where the first 3 words form the architecture, and the 4th word contains '-t' pattern

expts = [
    "resnet_layer_basic_t-5_extra",
    "resnet_layer_basic_t-10_extra",
    "vgg_conv_small_t-3_config",
    "vgg_conv_small_t-7_config",
    "bert_attention_head_t-2_setting",
]

ARGS = {
    "expts": expts
}
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_device Ran
ran on a synthesized input
danielmamay/grokking/grokking/training.py
code served (permissive licence) · get_code("c1c618e180dfc666")
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 device. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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

torch.manual_seed(0)

# The function get_device takes a single parameter 'device' which is a string
# It handles 'auto', 'cuda', 'mps', and 'cpu' device specifications
# We'll test with 'cpu' since we're on CPU-only environment

ARGS = {
    'device': 'cpu'
}
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 device — a device output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a device 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_loss_and_grads Ran
ran on a synthesized input
openai/grok/grok/measure.py
code served (permissive licence) · get_code("61f1915864fc4345")
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. 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
import numpy as np

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
    
    def _step(self, batch, batch_idx, train, grads):
        x, y = batch
        output = self.linear(x)
        loss = torch.mean((output - y) ** 2)
        
        if grads:
            loss.backward()
            grad_list = []
            for p in self.parameters():
                if p.grad is not None:
                    grad_list.append(p.grad.clone().flatten())
                else:
                    grad_list.append(torch.zeros_like(p).flatten())
            all_grads = torch.cat(grad_list)
            return loss.detach(), all_grads
        else:
            return loss.detach(), None

model = SimpleModel()

# Initialize x with flattened parameters
x_list = []
for p in model.parameters():
    x_list.append(p.data.cpu().numpy().flatten())
x = np.concatenate(x_list).astype(np.float64)

# Create data loader with small batches
data_loader = [
    (torch.randn(2, 4), torch.randn(2, 3)),
    (torch.randn(2, 4), torch.randn(2, 3))
]

ARGS = {
    'x': x,
    'model': model,
    'data_loader': data_loader
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is 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_sharpness Ran
ran on a synthesized input
openai/grok/grok/measure.py
code served (permissive licence) · get_code("739eddc60be1c197")
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 0, dtype family float, shape () 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
import numpy as np

torch.manual_seed(0)

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = nn.Linear(4, 3)
        self.linear2 = nn.Linear(3, 2)
    
    def forward(self, x):
        x = torch.relu(self.linear1(x))
        x = self.linear2(x)
        return x
    
    def _step(self, batch, batch_idx, train, grads):
        x, y = batch
        logits = self.forward(x)
        loss = nn.functional.mse_loss(logits, y)
        
        if grads:
            loss.backward()
            grad_list = []
            for p in self.parameters():
                if p.grad is not None:
                    grad_list.append(p.grad.data.view(-1))
                else:
                    grad_list.append(torch.zeros_like(p.data.view(-1)))
            grads_tensor = torch.cat(grad_list)
            return loss.detach(), grads_tensor.detach()
        else:
            return loss.detach(), None

model = SimpleModel()

data_loader = [
    (torch.randn(2, 4), torch.randn(2, 2)),
    (torch.randn(2, 4), torch.randn(2, 2)),
]

ARGS = {
    'data_loader': data_loader,
    'model': model,
    'subspace_dim': 5,
    'epsilon': 0.01,
    'maxiter': 2,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 0, dtype family float, shape () — shape () 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
gradfilter_ema Ran
ran; our contract draft was wrong, not the code
ironjr/grokfast/grokfast.py
code served (permissive licence) · get_code("b1849c24fbb69bfa")
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)

# Create a simple neural network module
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = nn.Linear(4, 3)
        self.linear2 = nn.Linear(3, 2)
    
    def forward(self, x):
        x = self.linear1(x)
        x = self.linear2(x)
        return x

# Instantiate the model
m = SimpleModel()

# Create a simple batch of data and compute gradients
x = torch.randn(2, 4)
target = torch.randint(0, 2, (2,))

# Forward pass
output = m(x)
loss = nn.CrossEntropyLoss()(output, target)

# Backward pass to populate gradients
loss.backward()

# Create initial grads dict (can be None, but let's test with a value too)
grads = {n: p.grad.data.detach().clone() for n, p in m.named_parameters() if p.requires_grad and p.grad is not None}

ARGS = {
    "m": m,
    "grads": grads,
    "alpha": 0.98,
    "lamb": 2.0
}
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
gradfilter_ma Ran
ran; our contract draft was wrong, not the code
ironjr/grokfast/grokfast.py
code served (permissive licence) · get_code("2617e928ebfb8675")
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 deque
from typing import Dict, Optional, Literal

torch.manual_seed(0)

# Create a simple model with parameters that have gradients
class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = nn.Linear(4, 3)
        self.linear2 = nn.Linear(3, 2)
    
    def forward(self, x):
        x = self.linear1(x)
        x = self.linear2(x)
        return x

# Build the model
m = SimpleModel()

# Create some dummy input and compute gradients
x = torch.randn(2, 4)
y = torch.randn(2, 2)
output = m(x)
loss = ((output - y) ** 2).sum()
loss.backward()

# Create grads dict with deques containing gradient tensors
window_size = 100
grads = {n: deque(maxlen=window_size) for n, p in m.named_parameters() if p.requires_grad and p.grad is not None}

# Populate with initial gradient values
for n, p in m.named_parameters():
    if p.requires_grad and p.grad is not None:
        grads[n].append(p.grad.data.detach().clone())

ARGS = {
    'm': m,
    'grads': grads,
    'window_size': 100,
    'lamb': 5.0,
    'filter_type': 'mean',
    'warmup': False,
    'trigger': 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
load_metric_data Ran
ran on a synthesized input
openai/grok/grok/visualization.py
code served (permissive licence) · get_code("7564745b254db80c")
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 os
import tempfile
import csv
from pathlib import Path

torch.manual_seed(0)

# Create a temporary directory structure with mock data
temp_dir = tempfile.mkdtemp()

# Create experiment subdirectories with metrics.csv files
expt_names = [
    "arch_v1_base_t-10",
    "arch_v1_base_t-20",
    "arch_v2_wide_t-15",
]

for expt_name in expt_names:
    expt_path = os.path.join(temp_dir, expt_name, "default", "version_0")
    os.makedirs(expt_path, exist_ok=True)
    
    # Create a minimal metrics.csv
    csv_path = os.path.join(expt_path, "metrics.csv")
    with open(csv_path, 'w', newline='') as f:
        writer = csv.DictWriter(f, fieldnames=[
            'learning_rate', 'train_loss', 'train_accuracy',
            'val_loss', 'val_accuracy', 'epoch'
        ])
        writer.writeheader()
        
        # Write 100000 rows of dummy data to satisfy epochs=100000
        for epoch in range(100000):
            writer.writerow({
                'learning_rate': '0.001',
                'train_loss': '0.5',
                'train_accuracy': '0.95',
                'val_loss': '0.6',
                'val_accuracy': '0.94',
                'epoch': str(epoch)
            })

ARGS = {
    'data_dir': temp_dir,
    'epochs': 100000,
    'load_partial_data': False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, tqdm — 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
most_interesting Ran
ran on a synthesized input
openai/grok/grok/visualization.py
code served (permissive licence) · get_code("ecac6edaeb33872c")
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

torch.manual_seed(0)

# Construct metric_data dict with architecture keys
# Each architecture has 'T', 'val_accuracy', and 'val_loss' tensors
metric_data = {
    'arch1': {
        'T': torch.tensor([1, 2, 3, 4]),
        'val_accuracy': torch.tensor([[92.0, 94.0, 96.0, 97.0],
                                      [91.0, 93.0, 95.0, 98.0]]),
        'val_loss': torch.tensor([[0.5, 0.3, 0.2, 0.1],
                                  [0.6, 0.4, 0.25, 0.08]])
    },
    'arch2': {
        'T': torch.tensor([1, 2, 3, 4]),
        'val_accuracy': torch.tensor([[88.0, 90.0, 92.0, 94.0],
                                      [89.0, 91.0, 93.0, 95.5]]),
        'val_loss': torch.tensor([[0.8, 0.6, 0.4, 0.3],
                                  [0.7, 0.5, 0.35, 0.2]])
    }
}

ARGS = {
    'metric_data': metric_data
}
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
norm Ran
ran, honoured the contract we drafted
openai/grok/grok/metrics.py
code served (permissive licence) · get_code("458780ffac6c9c58")
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 float, ndim 0, dtype family float, shape () 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)

class SimpleModule(nn.Module):
    def __init__(self):
        super().__init__()
        # Create a weight matrix with distinct dimensions (4 output units, 3 input features)
        self.weight = nn.Parameter(torch.randn(4, 3))

module = SimpleModule()
init_module = SimpleModule()

ARGS = {
    "module": module,
    "init_module": init_module,
    "p": 2,
    "q": 2
}
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 float, ndim 0, dtype family float, shape () — shape () 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
op_norm Ran
ran on a synthesized input
openai/grok/grok/metrics.py
code served (permissive licence) · get_code("7a0e73ebc4aab88f")
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 float, ndim 0, dtype family float, shape () 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)

class SimpleModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.weight = nn.Parameter(torch.randn(4, 3))
    
    def forward(self, x):
        return x @ self.weight.t()

module = SimpleModule()
init_module = SimpleModule()
p = float('Inf')

ARGS = {
    'module': module,
    'init_module': init_module,
    'p': p
}
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 float, ndim 0, dtype family float, shape () — shape () 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
get_data_loaders Not yet run
unverified (harvested, no recorded run)
danielmamay/grokking/grokking/data.py
code served (permissive licence) · get_code("ab942b3fd9d96975")
How it ran

no invocation recorded.

get_weights Not yet run
unverified (harvested, no recorded run)
openai/grok/grok/measure.py
code served (permissive licence) · get_code("269ed6282c593f2f")
How it ran

no invocation recorded.

operation_mod_p_data Not yet run
unverified (harvested, no recorded run)
danielmamay/grokking/grokking/data.py
code served (permissive licence) · get_code("f02ef440c5bcd946")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 13 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
norm 2 other papers
2 dated earlier
1706.08947 (2017-06)
Exploring Generalization in Deep Learning
1709.01953 (2017-09)
Implicit Regularization in Deep Learning

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 2026-11-21.

Repositories linked to this paper

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

Abstract

In this paper we propose to study generalization of neural networks on small algorithmically generated datasets. In this setting, questions about data efficiency, memorization, generalization, and speed of learning can be studied in great detail. In some situations we show that neural networks learn through a process of "grokking" a pattern in the data, improving generalization performance from random chance level to perfect generalization, and that this improvement in generalization can happen well past the point of overfitting. We also study generalization as a function of dataset size and find that smaller datasets require increasing amounts of optimization for generalization. We argue that these datasets provide a fertile ground for studying a poorly understood aspect of deep learning: generalization of overparametrized neural networks beyond memorization of the finite training dataset.

For agents

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

get_harvested_code_for_paper("2201.02177")
get_code_for_paper("2201.02177")
have("2201.02177")

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

curl https://syntology.ai/api/ran/2201.02177.json

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

[![Syntology run record](https://syntology.ai/api/ran/2201.02177.svg)](https://syntology.ai/paper/2201.02177)

Connect an agent — have() is free.