Orionfold/Kepler-bench

Kepler astro-bench v0.1

The benchmark behind Orionfold/Kepler-GGUF — a verifier-checked set of astrodynamics and quantitative-astrophysics word problems, each with a single numeric gold answer and a programmatic verifier that doubles as a reinforcement-learning reward.

What's here

File	Rows	Purpose
pool.jsonl	120	Training / selection pool — 16 formula families (9 orbital, 7 astrophysics), 3 difficulty tiers.
heldout.jsonl	44	External curveball held-out — different seeds + hand-curated edge cases, disjoint from the pool. The number on the model card is measured here.
verifier.py	—	astro_numeric_match(...) — the scorer.
units.py	—	SI-unit parsing/normalization used by the verifier.

Row schema

{
  "task_id": "astro-orb-leo_period-0000",
  "topic": "orbital_mechanics",
  "subtopic": "leo_period",
  "tier": 2,
  "prompt": "A satellite orbits at altitude h = 1,030 km ... Give your final answer as \\boxed{value unit}.",
  "answer": "105.6 min",
  "gold_value_si": 6336.46,
  "gold_unit": "s",
  "rel_tol": 0.02,
  "hand_curated": false,
  "params": {"h_km": 1030}
}

All physical constants are given in the prompt — the task tests reasoning, not memorization. The expected answer is a single \boxed{value unit}.

The verifier is the reward

astro_numeric_match extracts the \boxed{} answer, normalizes units to SI, and checks the value against the gold within a per-row relative tolerance (default ±2%). It returns a binary score, so it plugs directly into an RLVR loop as the reward — the same scorer used to build Kepler's SFT corpus, to gate the SFT checkpoint, and to run the head-to-head comparison.

from verifier import astro_numeric_match  # needs units.py alongside

reward = astro_numeric_match(
    completion=model_output,        # the model's full text, containing \boxed{...}
    expected="105.6 min",           # the row's "answer" field
    rel_tolerance=0.02,             # the row's "rel_tol" field
)  # -> 1.0 if correct within tolerance, else 0.0

Known coverage gaps

Honest about its weak spots: the families hohmann_transfer (two-burn transfers) and altitude_from_period (inverse Kepler) are the hardest rows and where models — including Kepler — most often miss. Treat them as the frontier of this benchmark.

Methods

Full construction + measurement protocol: The Gate Before the GPU — Deciding SFT vs RL vs RLVR Before You Spend the Run.

---

Published by Orionfold LLC · orionfold.com · Methods at ainative.business/field-notes.