No checkpoint selected
Choose a checkpoint button to inspect its receipts.
Nothing is selected merely because it is later or numerically larger.
A small instrument for accompanied seeing
See the path.
Do not mistake it for permission.
All 27 PIQA reports in one pinned Pythia directory, plus four richer receipts chosen after their values were known. Exact sources. No winner, no “latest” default, and no action inferred from a display.
Observation
Question
Authority
Frozen review snapshot
Loading reviewed source timestamp · no automatic refresh · manually recheck before treating it as current
All 27 reports in the pinned directory
The reported path, without a line
PIQA normalized accuracy · zero-shot
Evidence unavailable.
The pinned bundle did not pass its checks.
Every dot is one upstream report; error bars are reported standard errors. The vertical scale is the full 0–100%. Empty space is not interpolated.
all 27 reports 4 detail receipts
Pythia has 154 released weight-checkpoint positions; this pinned
directory has only 27 zero-shot reports, all shown here. The other
127 positions have no report in this directory. PIQA was selected
after outcomes were available from 87 tasks common to all 27 reports;
86 tasks are not shown. Four reports were also chosen after outcomes
for richer inspection, while 23 have source receipts only.
Representativeness is not claimed. Early reports overlap on the
honest full-run horizontal scale, so the labelled inset spreads
steps 0–1,000 logarithmically without moving them vertically or
adding jitter. Reported run configuration differs, including batch
size, device, and an early-only use_accelerate flag.
Exact harness revision, sample count, evaluated artifact identity,
and comparability remain not proven.
Read every one of the 27 reported values and exact source receipts
| Report | Derived tokens | Value ± stderr | Receipt in this release | Pinned source |
|---|
Four post-outcome-selected detail receipts
Four lanes, never collapsed
What kind of statement is this?
Observed
A value or artifact reference exists in an exact, pinned source.
Derived
Arithmetic is shown. Here, tokens seen equal step × 2,097,152.
Interpreted
A bounded claim names its rivals and can fail on held-out runs.
Authorized
A human or project steward permits one exact effect. Nothing here does.
The missing shape matters
Read what is not present.
A quiet omission can become hidden authority when an interface smooths it away. Pythia Paths keeps the gaps beside the evidence.
Not observed
Optimizer, scheduler, data-loader cursor, RNG state, exact harness revision, sample count.
Not proven
Metric comparability, metric-to-current-artifact binding, resumability, safety, deployment readiness, artifact-format equivalence.
Not inferred
Identity, experience, intention, preference, consent, destiny, or a hidden opposite.
Not implemented
Credentials, uploads, inference, jobs, persistence, telemetry, automatic refresh, or “continue.”
What changed beneath a name
Seed zero was not the same experiment.
Until February 2026, the repositories now named Pythia 14M and 31M held deduplicated-Pile models. PolyPythias’ archived seed-zero results point into those old histories, while seeds 1–9 used standard-Pile runs. Some published “seed” comparison therefore also changes data regime.
Historical seed 0deduplicated Pile
Seeds 1–9standard Pile
Name continuity did not establish artifact continuity.