Overall median throughput
Across the approved bounded probe rows.
Measured evidence
Training probe
Step counts, throughput ranges, parameter footprint, and pair-work reduction from bounded probes.
Recorded finding: Small controlled probes provide measurement signals suitable for engineering decisions and follow-up experiments.
On this page
Currency and next evidence gate
Decision signals
Key metrics and controls
The recorded values, with their controls and scope.
Minimum observed throughput
Minimum in the approved probe set.
Median pair-work reduction
Within the bounded probe configuration.
Trainable parameter share
125M of 520M parameters.
Source-linked visual evidence
What the evidence shows
Open any chart’s data table to inspect the exact values, or download the CSV.
Ordered source series
Probe throughput by sequence length
Minimum and median tokens per second across the approved sequence-length groups.
Finding: The 1024-token probe group reports higher minimum and median throughput than the 512-token group in this bounded run set.
512 tokens
1024 tokens
Connecting strokes aid comparison between published ordered points; they do not represent measurements between those points.
View source values
Scroll sideways if any columns are out of view.
| Point | Minimum | Median |
|---|---|---|
| 512 tokens | 304.6 tokens/s | 340.1 tokens/s |
| 1024 tokens | 417.0 tokens/s | 448.0 tokens/s |
Ordered source series
Pair-work reduction by sequence length
Median within-probe pair-work reduction for each approved sequence length.
Finding: The reported reduction is larger in the 1024-token probe group, within the current two-point scope.
512 tokens
1024 tokens
Connecting strokes aid comparison between published ordered points; they do not represent measurements between those points.
View source values
Scroll sideways if any columns are out of view.
| Point | Median reduction |
|---|---|
| 512 tokens | 1.28 × |
| 1024 tokens | 1.88 × |
Method
How to read this result
- The public summary groups 30 finite-loss optimizer-step rows across two sequence lengths.
- Probe throughput is an engineering signal and must not be generalized to production serving throughput.
What this result establishes
Probe-scale behavior is not a substitute for matched end-to-end production benchmarks.
- Local bounded training-throughput and trainability probe only.
- Parameter counts are normalized to M.
- No step-level rows, raw losses, config payloads, probe args, layer names, parameter names, block plans, memory snapshots, source paths, command names, script names, source internals, model weights, checkpoints, raw prompts, or raw datasets are bundled.
- No dense-baseline superiority, downstream quality, production-readiness, external benchmark reproduction, long-run training stability, or service-level claim.
- Pair-work reduction is an aggregate local probe signal, not an end-user capability or quality 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 JSONProvenance
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 receiptContinue reviewing
Related evidence
measuredreviewed
Historical- Observed
- The approved runs expose measurable resource behavior with source rows and bounded interpretation.
- Supports
- Public aggregate memory and smoke-loss guardrail curves across the approved experimental scales.
- Does not support
- These curves do not establish dense-baseline superiority or production performance.
- Next decisive evidence
- Repeat under a frozen current stack with matched controls before any broader comparison or performance use.
Decision relevance: research · diligence
Review evidencemeasuredpartial
Historical- Observed
- The training continuation path is instrumented and produces governed stability evidence.
- Supports
- Loss signals, stability controls, generation gates, reload behavior, and token-scale receipts.
- Does not support
- Generation quality and broad model capability remain under review and are not mature public claims.
- Next decisive evidence
- Run and seal a current exact-resume continuation with frozen evaluation and generation review.
Decision relevance: research
Review evidence