NeuroForge
Choose appearance
Appearance

Use your device setting or keep a preference on this browser.

Work with us
Navigate
See the 10-Day AI Fit TestChoose a path

Quick find

Navigate NeuroForge

Esc

Start typing to search public pages and evidence.

    Measured evidence

    Evidence curves

    Public aggregate memory and smoke-loss guardrail curves across the approved experimental scales.

    Recorded finding: The approved runs expose measurable resource behavior with source rows and bounded interpretation.

    On this page

    Decision signals

    Key metrics and controls

    Each value is mapped directly from the recorded summary artifact and retains its experimental, historical, or control context.

    47.9%

    Largest one-step RSS reduction

    Observed at 1000M; local one-step feasibility run.

    9,073 MiB

    Largest one-step RSS saving

    Difference between compared local peak-RSS measurements.

    50.1%

    Largest ten-step RSS reduction

    Observed at 1000M; local ten-step feasibility run.

    32.0%

    500M notebook eval-loss drop

    Separate approved notebook signal; not a dense-superiority result.

    Source-linked visual evidence

    Readable at a glance. Inspectable in detail.

    Charts are generated from normalized source fields. Every visualization includes its exact values in an accessible table.

    Ordered source series

    Peak resident memory by scale

    ERAIS and dense-control peak RSS in approved local one-step feasibility runs.

    Finding: The measured RSS gap increases across the four approved scale points, subject to the local-run boundary.

    Direct linkDownload CSV
    Peak resident memory by scaleERAIS and dense-control peak RSS in approved local one-step feasibility runs. Connecting lines show ordered observations and do not imply unmeasured intermediate results. The measured RSS gap increases across the four approved scale points, subject to the local-run boundary.05,11510,22915,34420,458MiBERAISDense control50M: ERAIS, 2,688 MiB100M: ERAIS, 3,080 MiB500M: ERAIS, 6,894 MiB1000M: ERAIS, 9,870 MiB50M: Dense control, 3,580 MiB100M: Dense control, 4,520 MiB500M: Dense control, 11,280 MiB1000M: Dense control, 18,943 MiB50M50M100M100M500M500M1000M1000M

    Connecting strokes aid comparison between published ordered points; they do not represent measurements between those points.

    View source values
    Peak resident memory by scale; values in MiB
    PointERAISDense control
    50M2,688 MiB3,580 MiB
    100M3,080 MiB4,520 MiB
    500M6,894 MiB11,280 MiB
    1000M9,870 MiB18,943 MiB

    Ordered source series

    Peak RSS reduction by scale

    Calculated percentage difference from the paired local peak-RSS values.

    Finding: The bounded runs report reductions from 24.9% at 50M to 47.9% at 1000M.

    Direct linkDownload CSV
    Peak RSS reduction by scaleCalculated percentage difference from the paired local peak-RSS values. Connecting lines show ordered observations and do not imply unmeasured intermediate results. The bounded runs report reductions from 24.9% at 50M to 47.9% at 1000M.0.012.925.938.851.7%Reduction50M: Reduction, 24.9 %100M: Reduction, 31.9 %500M: Reduction, 38.9 %1000M: Reduction, 47.9 %50M50M100M100M500M500M1000M1000M

    Connecting strokes aid comparison between published ordered points; they do not represent measurements between those points.

    View source values
    Peak RSS reduction by scale; values in %
    PointReduction
    50M24.9 %
    100M31.9 %
    500M38.9 %
    1000M47.9 %

    Source-linked scatter

    Memory saving by stored parameter scale

    Published parameter scale against the paired local peak-RSS saving for each approved one-step run.

    Finding: The four published observations show larger absolute RSS savings at larger stored-parameter labels; this remains bounded local evidence.

    Direct linkDownload CSV
    50M
    Stored parameters
    50 M parameters
    Peak RSS saving
    892 MiB saved
    100M
    Stored parameters
    100 M parameters
    Peak RSS saving
    1,440 MiB saved
    500M
    Stored parameters
    500 M parameters
    Peak RSS saving
    4,386 MiB saved
    1000M
    Stored parameters
    1,000 M parameters
    Peak RSS saving
    9,073 MiB saved
    Memory saving by stored parameter scalePublished parameter scale against the paired local peak-RSS saving for each approved one-step run. The four published observations show larger absolute RSS savings at larger stored-parameter labels; this remains bounded local evidence.02705408101,08002,4504,8997,3499,799M parametersMiB saved50M: 50 M parameters; 892 MiB saved50M100M: 100 M parameters; 1,440 MiB saved100M500M: 500 M parameters; 4,386 MiB saved500M1000M: 1,000 M parameters; 9,073 MiB saved1000M

    Points are published observations. Their placement does not claim measurements between or beyond the shown coordinates.

    View source values
    Memory saving by stored parameter scale
    PointStored parametersPeak RSS saving
    50M50 M parameters892 MiB saved
    100M100 M parameters1,440 MiB saved
    500M500 M parameters4,386 MiB saved
    1000M1,000 M parameters9,073 MiB saved

    Method

    How to interpret this pack

    1. The comparison uses paired local feasibility runs at 50M, 100M, 500M, and 1000M labels.
    2. Peak resident memory and smoke-loss values are published together so the resource result is not separated from its guardrail signal.
    Read the complete evidence method

    Claim boundaries

    These curves do not establish dense-baseline superiority or production performance.

    • Local evidence only.
    • 10-step desktop scale rows are feasibility/resource-shape evidence.
    • No final quality superiority claim.
    • No matched dense throughput superiority claim.
    • No external reproduction claim.

    Inspection and reuse

    Downloads and provenance

    Source data

    Machine-readable summary

    The recorded JSON artifact is staged from the governed public evidence pack during the build. Its currency notice determines whether it may be read as current.

    Open source JSON
    Peak resident memory by scale CSVPeak RSS reduction by scale CSVMemory saving by stored parameter scale CSV

    Provenance

    Build and evidence receipt

    Use the public build receipt to compare the deployed site source, evidence digest, claim posture, and inference release state.

    Open build receipt

    Continue reviewing

    Related evidence

    measuredreviewed
    Refresh required

    Training probe

    Observed
    Small controlled probes provide measurement signals suitable for engineering decisions and follow-up experiments.
    Supports
    Step counts, throughput ranges, parameter footprint, and pair-work reduction from bounded probes.
    Does not support
    Probe-scale behavior is not a substitute for matched end-to-end production benchmarks.
    Next decisive evidence
    Regenerate from the repaired exact artifact identity and close or retain the phase-41 contract-discovery failure.

    Decision relevance: research · engineering

    Review evidence
    measuredpartial
    Historical

    100M scale matrix

    Observed
    The scale matrix exposes a concrete engineering checkpoint with explicit dimensions and receipts.
    Supports
    Active parameters, evaluation loss, memory peak, and throughput-ratio results at the approved scale point.
    Does not support
    A single bounded scale matrix does not prove performance superiority across hardware or workloads.
    Next decisive evidence
    Repeat the matrix under one frozen current profile and matched control before current-stack use.

    Decision relevance: research · engineering

    Review evidence