---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- rl
- confidence-estimation
- calibration
- math
- code
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*.parquet
---

# rl-value-eval

Held-out **evaluation set** (774 prompts) for the Block-3 confidence / correctness-estimation
comparison. Pairs with [`YangyiYY/rl-value-confidence-train`](https://huggingface.co/datasets/YangyiYY/rl-value-confidence-train)
(training) and the nct-ppo actor/critic models. Each row is in the verl RL schema so the shared
grader scores it directly.

**Single `test` split** (the former `validation`+`test` splits are merged into one).

## Composition & subset annotation

Every row carries a fine-grained subset label in both `extra_info.sub_source` and a top-level
`sub_source` column:

| data_source | n | `sub_source` = |
|---|---|---|
| `taco` | 300 | TACO difficulty — EASY / MEDIUM / MEDIUM_HARD / HARD / VERY_HARD |
| `codeforces` | 300 | `rating_1300`, `rating_1400` |
| `math_box` (numina) | 98 | NuminaMath origin — cn_k12 / synthetic_math / orca_math / olympiads / gsm8k / synthetic_amc / aops_forum / amc_aime / math |
| `codecontests` | 76 | `rating_800` … `rating_1200` |

## Grader / ground_truth

`reward_model.ground_truth` is built for the **shared reference grader** (`reference/` in the
project repo): math → `{"answer": …}`; code → `input_output` JSON with the **full** hidden test
suites (CodeContests public+**private**+**generated**, Codeforces full suite, TACO raw
`input_output` incl. call-based `fn_name`), **no per-problem test-case cap**. Correctness =
last `\boxed{}` match (math) or all stdio/functional cases pass (code, 6s/case · 60s/problem).

## Schema

`data_source` · `prompt` (chat messages) · `ability` · `reward_model.ground_truth` · `sub_source` ·
`extra_info` (`id`, `rating`, `sub_source`).