SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2210.11693 · 2022

Amos: An Adam-style Optimizer with Adaptive Weight Decay towards Model-Oriented Scale

arXiv · PDF · Open in the Atlas

Code that ran

14 of 18 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; 0 honoured the contract we drafted, 0 violated it and 14 have no contract verdict; 0 failures of our instrument; 4 no recorded run), 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/2210.11693.json · badge: /api/ran/2210.11693.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
google-research/jestimator canonical (classifier label)
classifier basis not recorded on the link
14 ran · 4 no recorded run
FunctionStatusWhere it lives
amos Ran
ran on a synthesized input
google-research/jestimator/jestimator/amos.py
code served (permissive licence) · get_code("2bb6c564f3751916")
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 GradientTransformation. 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 jax
import jax.numpy as jnp
import optax
from typing import Tuple, Any

# Define eta_fn: maps (name, shape) to a scalar eta value
def eta_fn(name: Tuple[str, ...], shape: Tuple[int, ...]) -> float:
    return 0.1

# Define shape_fn (optional): maps (name, shape) to a reduced shape
def shape_fn(name: Tuple[str, ...], shape: Tuple[int, ...]) -> Tuple[int, ...]:
    # Return a shape with some dimensions reduced to 1
    if len(shape) > 1:
        return (1,) * len(shape)
    return shape

# Create a simple JAX pytree as params
# Using a nested dict structure that can be serialized/deserialized
params = {
    'layer1': jnp.array([[1.0, 2.0, 3.0, 4.0], [5.0, 6.0, 7.0, 8.0]]),
    'layer2': jnp.array([0.1, 0.2, 0.3, 0.4]),
    'nested': {
        'sublayer': jnp.array([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]])
    }
}

# Define learning rate as a float
learning_rate = 0.01

# Define learning rate as a schedule (callable)
def learning_rate_schedule(step):
    return 0.01 / (1.0 + 0.0001 * step)

ARGS = {
    'learning_rate': learning_rate,
    'eta_fn': eta_fn,
    'shape_fn': shape_fn,
    'beta': 0.999,
    'momentum': None,
    'clip_value': None,
    'extra_l2': 0.0,
    'd_coef': 0.25,
    'c_coef': 0.25,
    'epsilon': 1.0 / (1 << 125)
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, optax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type GradientTransformation — a GradientTransformation output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a GradientTransformation output the recorded ndim 1 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
count_dataset Ran
ran on a synthesized input
google-research/jestimator/jestimator/data_utils.py
code served (permissive licence) · get_code("6589ebccf58154e2")
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 1, dtype family int; 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 tensorflow as tf
import numpy as np

# Create a small tf.data.Dataset
# The function iterates through the dataset and accesses the first element
# We'll create a simple dataset of small tensors
data = tf.data.Dataset.from_tensor_slices(
    np.random.randn(10, 4).astype(np.float32)
)

ARGS = {
    "d": data,
    "batch_size": 3,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family int — 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
get_dataset_filenames Ran
ran on a synthesized input
google-research/jestimator/jestimator/data_utils.py
code served (permissive licence) · get_code("d8fb636561abd584")
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 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 tensorflow as tf
from typing import Union, Sequence, List

# Test case 1: Simple string pattern
pattern_1 = "dummy://test_data"

# Test case 2: List of patterns with mixed types
pattern_2 = ["dummy://data1", "seqio://task_name", "tfds://dataset/config"]

# Test case 3: tf.data.Dataset of strings
dataset_strings = tf.data.Dataset.from_tensor_slices([
    "dummy://path1",
    "tfds://test_dataset"
])

# Create arguments for the function
ARGS = {
    "pattern": "dummy://test_pattern",
    "do_glob": False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
tensorflow — 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
get_eta_fn Ran
ran on a synthesized input
google-research/jestimator/jestimator/models/bert_rpe/modeling.py
code served (permissive licence) · get_code("f35f7c250f470dec")
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 function. 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 dataclasses
import math
from typing import Tuple

@dataclasses.dataclass
class ModelConfig:
    """Config object."""
    max_length: int = 512
    hidden_size: int = 768
    num_heads: int = 12
    num_layers: int = 12
    relative_attention_max_distance: int = 128
    mlp_act: str = 'gelu'
    mlp_size: int = 3072
    initializer_range: float = 0.02
    attention_dropout_rate: float = 0.1
    hidden_dropout_rate: float = 0.1
    layer_norm_eps: float = 1e-12
    vocab_size: int = -1
    num_segments: int = 1
    num_labels: int = -1
    num_special_tokens: int = 0

config = ModelConfig(
    max_length=512,
    hidden_size=768,
    num_heads=12,
    num_layers=12,
    relative_attention_max_distance=128,
    mlp_act='gelu',
    mlp_size=3072,
    initializer_range=0.02,
    attention_dropout_rate=0.1,
    hidden_dropout_rate=0.1,
    layer_norm_eps=1e-12,
    vocab_size=30522,
    num_segments=2,
    num_labels=2,
    num_special_tokens=0
)

ARGS = {
    'config': config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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_shape_fn Ran
ran on a synthesized input
google-research/jestimator/jestimator/models/bert/modeling.py
code served (permissive licence) · get_code("e1fa28ec4113168f")
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 function. 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 get_shape_fn takes a config parameter
# Looking at the function body, it doesn't actually use the config (del config)
# So we can pass any object, including a dict or a simple class

class SimpleConfig:
    """Minimal config class."""
    pass

config = SimpleConfig()

ARGS = {
    '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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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_shape_fn Ran
ran on a synthesized input
google-research/jestimator/jestimator/models/bert_rpe/modeling.py
code served (permissive licence) · get_code("08fa84f42053f15a")
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 function. 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 typing import Tuple

torch.manual_seed(0)

Shape = Tuple[int, ...]

# Create a minimal config object
config = {
    'optimizer': 'amos',
    'learning_rate': 0.001,
}

ARGS = {
    '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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
last_evaluated_ckpt Ran
ran on a synthesized input
google-research/jestimator/jestimator/checkpoint_utils.py
code served (permissive licence) · get_code("93deb8b469412f3b")
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 str. 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 tempfile
import os
from tensorflow.io import gfile

# Create a temporary directory and file for testing
temp_dir = tempfile.mkdtemp()
last_eval_path = os.path.join(temp_dir, "last_eval.txt")

# Write a sample checkpoint path to the file
with gfile.GFile(last_eval_path, 'wb') as f:
    f.write(b"path/to/checkpoint-42")

ARGS = {
    "last_eval_path": last_eval_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
latest_ckpt_path Ran
ran on a synthesized input
google-research/jestimator/jestimator/checkpoint_utils.py
code served (permissive licence) · get_code("b13a2245fd60febd")
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 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 os
import tempfile
from unittest.mock import Mock, patch
from typing import Optional, Tuple

# Mock the JAX and Flax dependencies
class MockCheckpoints:
    @staticmethod
    def latest_checkpoint(model_dir, prefix='checkpoint_'):
        return None

class MockMultihostUtils:
    @staticmethod
    def sync_global_devices(name):
        pass

class MockGFile:
    @staticmethod
    def glob(pattern):
        return []
    
    @staticmethod
    def rmtree(path):
        pass

class MockJax:
    @staticmethod
    def process_index():
        return 0

class MockErrors:
    class NotFoundError(Exception):
        pass

# Create temporary directories for testing
temp_model_dir = tempfile.mkdtemp()
temp_init_ckpt_dir = tempfile.mkdtemp()
temp_init_ckpt_path = os.path.join(temp_init_ckpt_dir, "checkpoint_1")

# Create dummy checkpoint file
with open(temp_init_ckpt_path, 'w') as f:
    f.write("dummy checkpoint")

ARGS = {
    'model_dir': temp_model_dir,
    'init_ckpt_path': temp_init_ckpt_path,
    'prefix': 'checkpoint_'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, flax, jax, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, 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
normalize_loss_by_size Ran
ran on a synthesized input
google-research/jestimator/jestimator/modeling.py
code served (permissive licence) · get_code("3435635af154f27d")
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 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 jax.numpy as jnp

# The function takes two array-like parameters: loss and size
# Both need to be numeric values that can be converted to JAX arrays

# Create small numeric values for loss and size
loss_value = 16.0  # scalar loss value
size_value = 4.0   # scalar size value

# Alternatively, we could use arrays:
# loss_value = jnp.array([1.0, 2.0, 3.0])
# size_value = jnp.array([2.0, 2.0, 2.0])

ARGS = {
    "loss": loss_value,
    "size": size_value,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, 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
scale_by_amos Ran
ran on a synthesized input
google-research/jestimator/jestimator/amos.py
code served (permissive licence) · get_code("48d615d3ce949e18")
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 GradientTransformation. 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 jax
import jax.numpy as jnp
import optax
from typing import Tuple, Any

# Define a simple params structure (pytree)
params = {
    'layer1': jnp.ones((4, 3)),
    'layer2': jnp.ones((2,)),
}

# Define learning_rate as a float
learning_rate = 0.001

# Define eta_fn: takes (param_name: Tuple[str, ...], shape: Shape) and returns a scalar
def eta_fn(name: Tuple[str, ...], shape: Tuple[int, ...]) -> float:
    return 0.1

# Define shape_fn (optional): takes (param_name: Tuple[str, ...], shape: Shape) and returns a new shape
def shape_fn(name: Tuple[str, ...], shape: Tuple[int, ...]) -> Tuple[int, ...]:
    # For testing, reduce last dimension by half
    return shape

# Define other scalar parameters
beta = 0.999
extra_l2 = 0.0
d_coef = 0.25
c_coef = 0.25
epsilon = 1.0 / (1 << 125)

ARGS = {
    'learning_rate': learning_rate,
    'eta_fn': eta_fn,
    'shape_fn': shape_fn,
    'beta': beta,
    'extra_l2': extra_l2,
    'd_coef': d_coef,
    'c_coef': c_coef,
    'epsilon': epsilon,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
flax, jax, optax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type GradientTransformation — a GradientTransformation output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a GradientTransformation output the recorded ndim 1 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
sparse_xe_with_logits Ran
ran on a synthesized input
google-research/jestimator/jestimator/modeling.py
code served (permissive licence) · get_code("7aa1ce8753bf1f47")
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 ArrayImpl, 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 jax
import jax.numpy as jnp

# Set seed for reproducibility
jax.config.update('jax_default_prng_impl', 'threefry2x32')
key = jax.random.PRNGKey(0)

# Create small test tensors
batch_size = 2
num_labels = 4
sequence_length = 3

# labels: int tensor of shape (batch_size, sequence_length)
# Each element is a class index from 0 to num_labels-1
labels = jnp.array([[0, 1, 2], [3, 0, 1]], dtype=jnp.int32)

# logits: float tensor of shape (batch_size, sequence_length, num_labels)
# Random logits for each position
logits = jax.random.normal(key, shape=(batch_size, sequence_length, num_labels))

# mask: 0/1 float tensor, same shape as labels (batch_size, sequence_length)
mask = jnp.array([[1.0, 1.0, 0.0], [1.0, 1.0, 1.0]], dtype=jnp.float32)

ARGS = {
    'labels': labels,
    'logits': logits,
    'mask': mask,
    'normalized': False,
    'reduce_all': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, 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
to_attention_mask Ran
ran on a synthesized input
google-research/jestimator/jestimator/models/bert/modeling.py
code served (permissive licence) · get_code("31bc709a794648a0")
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 ArrayImpl, ndim 4, dtype family float, shape (2, 1, 1, 4) 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 jax.numpy as jnp
import numpy as np

# Create a simple input mask
# The function expects ArrayLike input (can be list, numpy array, or jax array)
# It will be converted to boolean and used to create attention mask
input_mask = np.array([
    [1, 1, 0, 1],
    [1, 0, 0, 1]
], dtype=np.int32)

ARGS = {
    "input_mask": input_mask
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 4, dtype family float, shape (2, 1, 1, 4) — shape (2, 1, 1, 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
transpose_dataset Ran
ran on a synthesized input
google-research/jestimator/jestimator/data_utils.py
code served (permissive licence) · get_code("2f530f6e8d717b5a")
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 _FlatMapDataset. 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 tensorflow as tf
import numpy as np

# Create a dataset of datasets
# The outer dataset will have size_d=3 elements
# Each element is itself a dataset with size_per_elem=4 elements
# Each inner element is a batch of size 2 with 5 features

def create_inner_dataset(batch_id):
    """Create a small inner dataset"""
    data = tf.data.Dataset.from_tensor_slices(
        np.random.randn(4, 5).astype(np.float32)
    )
    return data

# Create outer dataset by mapping over indices
outer_dataset = tf.data.Dataset.range(3).map(
    lambda i: create_inner_dataset(i)
)

size_d = 3          # number of inner datasets
size_per_elem = 4   # number of elements in each inner dataset
bs = 2              # batch size for final batching

ARGS = {
    "d": outer_dataset,
    "size_d": size_d,
    "size_per_elem": size_per_elem,
    "bs": bs
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, tensorflow — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type _FlatMapDataset — a _FlatMapDataset output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a _FlatMapDataset 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
unstack Ran
ran on a synthesized input
google-research/jestimator/jestimator/modeling.py
code served (permissive licence) · get_code("c1163fa4442ceef8")
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 list, ndim 3, dtype family int; 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 jax
import jax.numpy as jnp

# Create a small test tensor with distinct dimensions
x = jnp.arange(24).reshape(2, 3, 4)

# axis parameter - choose a valid axis to unstack along
axis = 1

ARGS = {
    "x": x,
    "axis": axis,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 3, dtype family int — 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
extract_axes Not yet run
unverified (harvested, no recorded run)
google-research/jestimator/jestimator/states.py
code served (permissive licence) · get_code("0cdb17a2cb7ed2ea")
How it ran

no invocation recorded.

get_eta_fn Not yet run
unverified (harvested, no recorded run)
google-research/jestimator/jestimator/models/bert/modeling.py
code served (permissive licence) · get_code("03eacb055758e4ce")
How it ran

no invocation recorded.

get_mode_heuristic Not yet run
unverified (harvested, no recorded run)
google-research/jestimator/jestimator/estimator.py
code served (permissive licence) · get_code("ae4f81e47ffe8ec4")
How it ran

no invocation recorded.

maybe_reduce_axis_names Not yet run
unverified (harvested, no recorded run)
google-research/jestimator/jestimator/amos_helper.py
code served (permissive licence) · get_code("8eb387dfd7170af5")
How it ran

no invocation recorded.

Also present in other papers' code

Of 18 samples lifted from this paper's linked repositories, 18 carry a function identity and 0 do not. Of the 18 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 18 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.

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

We present Amos, a stochastic gradient-based optimizer designed for training deep neural networks. It can be viewed as an Adam optimizer with theoretically supported, adaptive learning-rate decay and weight decay. A key insight behind Amos is that it leverages model-specific information to determine the initial learning-rate and decaying schedules. When used for pre-training BERT variants and T5, Amos consistently converges faster than the state-of-the-art settings of AdamW, achieving better validation loss within &lt;=70% training steps and time, while requiring &lt;=51% memory for slot variables. Our code is open-sourced at: https://github.com/google-research/jestimator

For agents

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

get_harvested_code_for_paper("2210.11693")
get_code_for_paper("2210.11693")
have("2210.11693")

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

curl https://syntology.ai/api/ran/2210.11693.json

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

[![Syntology run record](https://syntology.ai/api/ran/2210.11693.svg)](https://syntology.ai/paper/2210.11693)

Connect an agent — have() is free.