SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.19146 · NeurIPS · 2024

Resolving Discrepancies in Compute-Optimal Scaling of Language Models

Mitchell Wortsman, Yair Carmon, Ludwig Schmidt, Jenia Jitsev, Tomer Porian

arXiv · PDF · Open in the Atlas

Code that ran

10 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; 2 honoured the contract we drafted, 3 violated it and 0 have no contract verdict; 5 failures of our instrument; 6 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/2406.19146.json · badge: /api/ran/2406.19146.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
formll/resolving-scaling-law-discrepencies canonical (classifier label)
classifier basis: census
6 ran · 6 no recorded run
formll/resolving-scaling-law-discrepancies canonical (classifier label)
classifier basis: classifier_2025
3 ran · 0 no recorded run
copy not recorded community
not in the paper's declared repositories
1 ran · 0 no recorded run
FunctionStatusWhere it lives
add_zero Ran
ran; our contract draft was wrong, not the code
formll/resolving-scaling-law-discrepencies/paper_tables.py
code served (permissive licence) · get_code("6a004a1784aa8d01")
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 torch

def add_zero(x):
    if len(x.replace('.', '')) < 4:
        if '.' not in x:
            return x + '.0'
        return x + '0' * (4 - len(x.replace('.', '')))
    if len(x.replace('.', '')) == 5 and x[0] == '0':
        return x[:5]
    return x

ARGS = {
    'x': '3.14'
}
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 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
custom_huber_loss Ran
ran, violated the contract we drafted
formll/resolving-scaling-law-discrepancies/analysis.py
code served (permissive licence) · get_code("752782550c94c617")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned float64, 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 numpy as np

np.random.seed(0)

y_true = np.array([1.5, 2.3, 0.8, 1.2])
y_pred = np.array([1.4, 2.5, 0.9, 1.1])
delta = 0.001

ARGS = {
    'y_true': y_true,
    'y_pred': y_pred,
    'delta': delta
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, 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
fetch_flop Ran
ran, honoured the contract we drafted
formll/resolving-scaling-law-discrepancies/analysis.py
code served (permissive licence) · get_code("15bb4fddf111e4b7")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned DataFrame. 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 pandas as pd
import numpy as np

# Create a sample DataFrame with the required structure
# The function expects rows with loss_key containing Series with multi-index or regular index

# Create sample loss data for a few runs
np.random.seed(0)

# Create loss series with step indices
def create_loss_series(n_steps=100, start_step=10):
    steps = np.arange(start_step, start_step + n_steps)
    losses = 5.0 * np.exp(-steps / 50.0) + np.random.randn(n_steps) * 0.1
    return pd.Series(losses, index=steps)

# Build rows for the dataframe
rows = []
for i in range(3):
    loss_series = create_loss_series(n_steps=150, start_step=5)
    rows.append({
        'train/loss_smoothed': loss_series,
        'params': 1e6 * (i + 1),  # n_key
        'bs': 32,  # bs_key
        'seq_len': 2048,
        'warmup_tokens': 1e8,
        'flops_per_token': 6.0 * 1e6 * (i + 1),  # flop_per_token_key
        'lr': 1e-4
    })

# Add a row with empty loss to test that branch
rows.append({
    'train/loss_smoothed': pd.Series([], dtype=float),
    'params': 5e6,
    'bs': 32,
    'seq_len': 2048,
    'warmup_tokens': 1e8,
    'flops_per_token': 6e6,
    'lr': 1e-4
})

df = pd.DataFrame(rows)

# Choose a flop value that should be within the range of computed flops
# flop = step * seq_len * bs * flops_per_token
# For first row: step ~50, seq_len=2048, bs=32, flops_per_token=6e6
# flop ~ 50 * 2048 * 32 * 6e6 = 1.966e13
flop = 2e13

ARGS = {
    'df': df,
    'flop': flop,
    'loss_key': 'train/loss_smoothed',
    'warmup_remove_factor': 1e-12,
    'n_key': 'params',
    'seq_len': 2048,
    'bs_key': 'bs',
    'keep_bs_lr_keys': False,
    'flop_per_token_key': 'flops_per_token',
    'flop_tolerance': 0.1
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DataFrame — a DataFrame output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DataFrame output the recorded ndim 2 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_color Ran
ran; our contract draft was wrong, not the code
formll/resolving-scaling-law-discrepencies/utils.py
pointer only (licence: MIT); the stored copy get_code serves, from hadasah/scaling_laws, is pointer only (licence: NONE) · get_code("8d1a5df6af7bf845")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from hadasah/scaling_laws, 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_linestyle Ran
ran; our contract draft was wrong, not the code
formll/resolving-scaling-law-discrepencies/utils.py
code served (permissive licence) · get_code("99378ea8da320872")
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 torch

ARGS = {
    'i': 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 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
get_marker Ran
ran; our contract draft was wrong, not the code
formll/resolving-scaling-law-discrepencies/utils.py
code served (permissive licence) · get_code("9bcb697d9afad429")
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 torch

ARGS = {
    'i': 5
}
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 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
huber_loss_objective Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
formll/resolving-scaling-law-discrepancies/analysis.py
code served (permissive licence) · get_code("cd1ae572668e4ce4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) and returned float64, 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 numpy as np

np.random.seed(0)

# params is a tuple/list of 3 elements: a, e, alpha
# These should be scalars or arrays that can be used in arithmetic operations
a = 2.0
e = 1.5
alpha = 0.5
params = [a, e, alpha]

# F is an array used in np.log(F), so must be positive
# Shape should be distinct, let's use 5 elements
F = np.array([1.2, 2.3, 3.4, 4.5, 5.6])

# losses is an array used in np.log(losses), so must be positive
# Should have same shape as F (since predictions will have same shape as F)
losses = np.array([1.1, 2.2, 3.3, 4.4, 5.5])

ARGS = {
    'params': params,
    'F': F,
    'losses': losses
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float64, 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
maybe_get_item Ran
ran, violated the contract we drafted
this paper's copy was not recorded; identical code first harvested from hadasah/scaling_laws
pointer only; the stored copy get_code serves, from hadasah/scaling_laws, is pointer only (licence: NONE) · get_code("a6d9e09c75556988")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

precise_flops_per_token_chinchilla Ran
ran, violated the contract we drafted
formll/resolving-scaling-law-discrepencies/data.py
pointer only (licence: MIT); the stored copy get_code serves, from hadasah/scaling_laws, is pointer only (licence: NONE) · get_code("fc9bc7439b7fd07b")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from hadasah/scaling_laws, 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.

precise_param_count_open_lm Ran
ran, honoured the contract we drafted
formll/resolving-scaling-law-discrepencies/data.py
pointer only (licence: MIT); the stored copy get_code serves, from hadasah/scaling_laws, is pointer only (licence: NONE) · get_code("e20ba47c892c9025")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from hadasah/scaling_laws, 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.

draw_fit Not yet run
unverified (harvested, no recorded run)
formll/resolving-scaling-law-discrepencies/plotting.py
code served (permissive licence) · get_code("4fa5926c92763b45")
How it ran

no invocation recorded.

extract_CI Not yet run
unverified (harvested, no recorded run)
formll/resolving-scaling-law-discrepencies/paper_tables.py
code served (permissive licence) · get_code("4c878b8671db62ea")
How it ran

no invocation recorded.

extract_label Not yet run
unverified (harvested, no recorded run)
formll/resolving-scaling-law-discrepencies/paper_tables.py
code served (permissive licence) · get_code("9370894d3db38de0")
How it ran

no invocation recorded.

isoflop_curves_plot Not yet run
unverified (harvested, no recorded run)
formll/resolving-scaling-law-discrepencies/plotting.py
code served (permissive licence) · get_code("51aa561ef23bceaa")
How it ran

no invocation recorded.

opt_param_vs_compute_plot Not yet run
unverified (harvested, no recorded run)
formll/resolving-scaling-law-discrepencies/plotting.py
code served (permissive licence) · get_code("3ae2bda49e8d6539")
How it ran

no invocation recorded.

proportional_sliding_window_filter Not yet run
unverified (harvested, no recorded run)
formll/resolving-scaling-law-discrepencies/data.py
code served (permissive licence) · get_code("b8859ce442ad4d05")
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, 5 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 0 the same month, 5 later) and 11 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
get_color 2 other papers
1 dated earlier, 1 later
2404.10102 (2024-04)
Chinchilla Scaling: A replication attempt
2502.18969 (2025-02)
(Mis)Fitting: A Survey of Scaling Laws
maybe_get_item 2 other papers
1 dated earlier, 1 later
2404.10102 (2024-04)
Chinchilla Scaling: A replication attempt
2502.18969 (2025-02)
(Mis)Fitting: A Survey of Scaling Laws
custom_huber_loss 1 other paper
1 later
2502.18969 (2025-02)
(Mis)Fitting: A Survey of Scaling Laws
precise_flops_per_token_chinchilla 1 other paper
1 later
2502.18969 (2025-02)
(Mis)Fitting: A Survey of Scaling Laws
precise_param_count_open_lm 1 other paper
1 later
2502.18969 (2025-02)
(Mis)Fitting: A Survey of Scaling Laws

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

and Hoffmann et al. [25] developed influential scaling laws for the optimal model size as a function of the compute budget, but these laws yield substantially different predictions. We explain the discrepancy by reproducing the Kaplan et al. scaling law on two datasets (OpenWebText2 and RefinedWeb) and identifying three factors causing the difference: last layer computational cost, warmup duration, and scale-dependent optimizer tuning. With these factors corrected, we obtain excellent agreement with the Hoffmann et al. (i.e., "Chinchilla") scaling law. Counter to a hypothesis implied in Hoffmann et al. [25], we find that careful learning rate decay is not essential for the validity of their scaling law. As a secondary result, we derive scaling laws for the optimal learning rate and batch size, finding that tuning the AdamW β 2 parameter is essential at lower batch sizes.

For agents

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

get_harvested_code_for_paper("2406.19146")
get_code_for_paper("2406.19146")
have("2406.19146")

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

curl https://syntology.ai/api/ran/2406.19146.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.19146.svg)](https://syntology.ai/paper/2406.19146)

Connect an agent — have() is free.