FineProofs Rollout Distribution

Complete collection

Grade and value distributions

Prefix value Mean normalized rubric grade over 32 continuations

Individual continuation grades

All normalized rubric scores

n = 3,312,576

Credit composition

Exact zero, partial, and full credit

Prefix mean grades

One empirical value per collected prefix

Median -

Value across reasoning depth

Mean and interquartile range within normalized-depth deciles

All traces Endpoint grade at least 0.5 Endpoint grade below 0.5

Trajectory

-

Problem

Mean grade by prefix

Empirical quartiles across 32 continuations; diamond marks the base endpoint grade

-

Base model output

Reasoning trace

Selected prefix

Range: -
Steps: -
Characters: -

Loading reasoning trace

Ungraded tail Not a collected prefix

Selected prefix

-

Continuation grade distribution

32 samples

Run semantics

How to read the collection

01

Prefix value

Every partial point is the arithmetic mean of 32 normalized rubric grades from fresh continuations of that exact response prefix. It is not a binary success probability.

02

Trajectory endpoint

The diamond is the original base response's continuous rubric grade. The downstream PRM dataset separately thresholds eligible endpoints at 0.5 for a hard yes/no label.

03

Empty prefix

Each problem contributes two selected base traces, so its empty prefix appears once per trace in the raw branch set. The 103,518 rows correspond to 99,354 unique semantic problem-prefix inputs.

04

PRM filtering

The downstream training build removes duplicate empty prefixes, invalid endpoints, and overlength prompts while preserving provenance counts.