Datasets:

HerrHruby
/

soda_fa0_luna_sample

license: other
language:
  - en
task_categories:
  - text-generation
tags:
  - math
  - proofs
  - reasoning
  - rollouts
  - sample
  - example
  - llm-judge
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

SODA FA-0 + Luna — 10-row format sample

A tiny, public 10-row sample to illustrate the data format of the (private) full dataset HerrHruby/soda_fa0_luna. Use it to understand the schema before requesting/loading the full set — it is not meant for training or evaluation.

Each row is one FA-0 rollout (a model's direct final answer, no exploration) on a SODA-2026 proof problem, graded by GPT-5.6-Luna against a per-problem rubric.

⚠️ Long text fields are truncated in this sample. thinking, answer, luna_rubric, luna_explanation, and problem are cut (with a … [truncated for sample — full text (N chars) …] marker) so the file stays small and readable. The full dataset has them untruncated — real thinking/answer can be tens of thousands of characters (and when a rollout never closes </think>, the whole runaway generation lands in answer).

These 10 rows cover

both categories — 6 proof_writing, 4 proof_strategy
both rubric maxes — luna_max_score 7 and 8
8 graded (luna_concluded=True) + 2 not — one no_think_close auto-zero and one score-parse failure (rubric text present, numeric score null)
a reward spread: luna_reward ∈ {0.0 (incl. a graded zero), 0.357, 0.375, 0.429, 0.857, 1.0}

Columns

column	type	description
`problem_id`	str	`train:<i>` — stable id (8 samples share it in the full set)
`sample_idx`	int	FA-0 sample index
`source`	str	`SODA`
`category`	str	`proof_writing` or `proof_strategy`
`problem`	str	the problem statement
`thinking`	str	model reasoning (the `<think>` content)
`answer`	str	the final answer / written proof (after `</think>`)
`luna_score`	int?	rubric points earned (null on parse-failure)
`luna_max_score`	int?	rubric max, 7 or 8 (null if auto-zeroed)
`luna_reward`	float	normalized `[0,1]` = `luna_score / luna_max_score`, else `0.0`
`luna_concluded`	bool	True if the answer closed `</think>` and got a real rubric grade
`luna_auto_reason`	str?	`no_think_close` when auto-zeroed (else null)
`luna_rubric`	str	Luna's rubric-by-rubric grading (`raw_judge`)
`luna_explanation`	str	Luna's free-form explanation (distinct from the rubric)
`problem_luna_mean`	float	mean `luna_reward` over the problem's 8 samples (full set)
`problem_oss_mean`	float	per-problem gpt-oss-120b mean (gated selection into the Luna pool)
`n_layers_completed`	int	0 (FA-0)
`termination_reason`	str	e.g. `max_layers`
`wall_time_s`	float	rollout wall-clock

Load

from datasets import load_dataset
ds = load_dataset("HerrHruby/soda_fa0_luna_sample", split="train")   # public, 10 rows
print(ds[0])