SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2303.11257 · 2023

Unit Scaling: Out-of-the-Box Low-Precision Training

arXiv · PDF · Open in the Atlas

Code that ran

12 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; 0 honoured the contract we drafted, 0 violated it and 12 have no contract verdict; 0 failures of our instrument; 2 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/2303.11257.json · badge: /api/ran/2303.11257.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
graphcore-research/unit-scaling-demo canonical (classifier label)
classifier basis not recorded on the link
12 ran · 2 no recorded run
FunctionStatusWhere it lives
batched_gather Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/layers.py
code served (permissive licence) · get_code("2f8c9dc17d697ee0")
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 EagerTensor, ndim 2, dtype family float, shape (2, 3) 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 tensorflow as tf
import numpy as np

# Set random seeds for reproducibility
tf.random.set_seed(0)
np.random.seed(0)

# Create small test tensors
# tables: shape (batch, seq_len, vocab_size) - 3D tensor
# We'll use batch=2, seq_len=3, vocab_size=4
tables = tf.constant(np.random.randn(2, 3, 4).astype(np.float32))

# indices: shape (batch, seq_len) - 2D tensor with values in range [0, vocab_size)
# Same batch and seq_len as tables, but one fewer dimension
indices = tf.constant(np.random.randint(0, 4, size=(2, 3)), dtype=tf.int32)

ARGS = {
    "tables": tables,
    "indices": indices,
}
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 EagerTensor, ndim 2, dtype family float, shape (2, 3) — shape (2, 3) 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
conv1d Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/uscale/ops.py
code served (permissive licence) · get_code("922e7a65963e6b06")
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 EagerTensor, ndim 3, dtype family float, shape (2, 8, 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 tensorflow as tf
import numpy as np

# Set seed for reproducibility
tf.random.set_seed(0)
np.random.seed(0)

# Create small test tensors
batch_size = 2
input_length = 8
channels_in = 4
filter_width = 3
channels_out = 6

# Input tensor: (batch_size, input_length, channels_in)
input_tensor = tf.random.normal([batch_size, input_length, channels_in], dtype=tf.float32)

# Filters tensor: (filter_width, channels_in, channels_out)
filters_tensor = tf.random.normal([filter_width, channels_in, channels_out], dtype=tf.float32)

ARGS = {
    'input': input_tensor,
    'filters': filters_tensor,
    'stride': 1,
    'padding': 'SAME'
}
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 EagerTensor, ndim 3, dtype family float, shape (2, 8, 6) — shape (2, 8, 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_character Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/datasets.py
code served (permissive licence) · get_code("05df62f6648200c1")
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 Data. 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 json
from pathlib import Path
import numpy as np

# Create a temporary directory structure for the test
tmpdir = tempfile.TemporaryDirectory()
root = Path(tmpdir.name)

# Create vocab.json
vocab_list = ['a', 'b', 'c', 'd', 'e']
(root / 'vocab.json').write_text(json.dumps(vocab_list))

# Create train.txt with simple character data
train_text = "abcdeabcdeabcde"
(root / 'train.txt').write_text(train_text)

# Create valid.txt with simple character data
valid_text = "bcdea"
(root / 'valid.txt').write_text(valid_text)

ARGS = {
    'root': root,
    'train': 'train.txt',
    'valid': 'valid.txt'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Data — a Data output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Data 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
pointwise Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/uscale/ops.py
code served (permissive licence) · get_code("73c30a2cc2fee462")
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 EagerTensor, ndim 2, dtype family float, shape (2, 2) 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 tensorflow as tf
import numpy as np

# Create small tensors for testing
inputs = tf.constant([[1.0, 2.0, 3.0, 4.0],
                      [5.0, 6.0, 7.0, 8.0]], dtype=tf.float32)

weights = tf.constant([[0.1, 0.2],
                       [0.3, 0.4],
                       [0.5, 0.6],
                       [0.7, 0.8]], dtype=tf.float32)

ARGS = {
    'inputs': inputs,
    'weights': weights,
    'scale_for': 'both'
}
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 EagerTensor, ndim 2, dtype family float, shape (2, 2) — shape (2, 2) 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
printing Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/uscale/utility.py
code served (permissive licence) · get_code("8cae868db43ba55a")
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 Bind. 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 sys
import numpy as np
import tensorflow as tf

# Create a small TensorFlow tensor for testing
test_tensor = tf.constant([[1.0, 2.0, 3.0, 4.0], 
                           [5.0, 6.0, 7.0, 8.0]], dtype=tf.float32)

# Define a custom summary function that takes a tensor and returns a scalar
def custom_summary(x):
    return float(tf.reduce_mean(x))

# Set up the arguments dictionary
ARGS = {
    "name": "test_layer",
    "summary": np.std
}
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 Bind — a Bind output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Bind 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
scaling Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/uscale/ops.py
code served (permissive licence) · get_code("55834bdb013ab248")
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 Bind. 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

ARGS = {
    "forward": 2.0,
    "backward": 3.0,
}
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 Bind — a Bind output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Bind 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
sinusoid_embedding Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/layers.py
code served (permissive licence) · get_code("288d2d846f7173a7")
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 ndarray, ndim 2, dtype family float, shape (16, 8) 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 numpy as np

# Construct arguments for sinusoid_embedding
ARGS = {
    "sequence_length": 16,
    "frequencies": 8,
    "max_period": 10000,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 2, dtype family float, shape (16, 8) — shape (16, 8) 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
softmax_cross_entropy Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/layers.py
code served (permissive licence) · get_code("d380f15ca83de673")
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 tensorflow as tf

# Set seed for reproducibility
tf.random.set_seed(0)

# Create small tensors with distinct dimensions
batch_size = 2
sequence_length = 3
vocab_size = 4

# scores: [batch, seq_len, vocab_size] - logits for each token
scores = tf.random.normal([batch_size, sequence_length, vocab_size], dtype=tf.float32)

# ids: [batch, seq_len] - target token ids, values in range [0, vocab_size)
ids = tf.constant([[1, 2, 0], [3, 1, 2]], dtype=tf.int32)

# mask: [batch, seq_len] - binary mask indicating which positions to include in loss
mask = tf.constant([[1.0, 1.0, 1.0], [1.0, 0.0, 1.0]], dtype=tf.float32)

ARGS = {
    "scores": scores,
    "ids": ids,
    "mask": mask,
}
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 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
split_seed Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/pedal/utility.py
code served (permissive licence) · get_code("9e76a013a5849311")
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 numpy as np

# split_seed takes two parameters:
# - seed: int - a random seed value
# - n: int - number of seeds to split into

ARGS = {
    "seed": 42,
    "n": 5,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — 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
to_ids Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/datasets.py
code served (permissive licence) · get_code("c273c7f894008b87")
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 ndarray, ndim 1, dtype family int, shape (5,) 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 numpy as np

# Build small valid arguments for to_ids function
data = "hello"
vocab = ("h", "e", "l", "o")
dtype = np.uint8

ARGS = {
    "data": data,
    "vocab": vocab,
    "dtype": dtype,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family int, shape (5,) — shape (5,) 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_jsonable Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/pedal/utility.py
code served (permissive licence) · get_code("7b5e4d5d64c189c4")
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 datetime
from pathlib import Path
import numpy as np

# Test cases for to_jsonable function
# The function handles Path, np.ndarray, np.number, and datetime.date objects

ARGS = {
    "obj": Path("/home/user/documents/file.txt")
}

# Additional test cases can be created by calling the function with different object types:
# - np.ndarray example: np.array([1, 2, 3])
# - np.number example: np.int64(42)
# - datetime.date example: datetime.date(2023, 10, 15)
# - Invalid type example: object() would raise TypeError
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — 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
to_str Ran
ran on a synthesized input
graphcore-research/unit-scaling-demo/scmm/datasets.py
code served (permissive licence) · get_code("e30037d5501d3b17")
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 numpy as np

# Create a small vocabulary as a tuple of strings
vocab = ('a', 'b', 'c', 'd', 'e', 'f')

# Create a small numpy array of term indices
terms = np.array([0, 1, 2, 3, 1, 0])

ARGS = {
    'terms': terms,
    'vocab': vocab,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — 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
context Not yet run
unverified (harvested, no recorded run)
graphcore-research/unit-scaling-demo/scmm/pedal/xpu.py
code served (permissive licence) · get_code("b910812684c6f413")
How it ran

no invocation recorded.

remove_keys Not yet run
unverified (harvested, no recorded run)
graphcore-research/unit-scaling-demo/scmm/pedal/utility.py
code served (permissive licence) · get_code("f525cde07e4a95c4")
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, 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 14 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 unit scaling, a paradigm for designing deep learning models that simplifies the use of low-precision number formats. Training in FP16 or the recently proposed FP8 formats offers substantial efficiency gains, but can lack sufficient range for out-of-the-box training. Unit scaling addresses this by introducing a principled approach to model numerics: seeking unit variance of all weights, activations and gradients at initialisation. Unlike alternative methods, this approach neither requires multiple training runs to find a suitable scale nor has significant computational overhead. We demonstrate the efficacy of unit scaling across a range of models and optimisers. We further show that existing models can be adapted to be unit-scaled, training BERT-Large in FP16 and then FP8 with no degradation in accuracy.

For agents

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

get_harvested_code_for_paper("2303.11257")
get_code_for_paper("2303.11257")
have("2303.11257")

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

curl https://syntology.ai/api/ran/2303.11257.json

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

[![Syntology run record](https://syntology.ai/api/ran/2303.11257.svg)](https://syntology.ai/paper/2303.11257)

Connect an agent — have() is free.