amphora/math-intuition-reasoning-traces

math-intuition reasoning traces

Full chain-of-thought traces from 7 reasoning models on the same 4,020 problems, graded by each problem family's own verifier.

Questions come from amphora/math-intuition-20260908-402-easy-10 — 402 arXiv-derived problem families x 10 seeds, easy preset. Every row here refers to an id in that dataset, so prompts and the instance cache can be joined from it.

Generation settings

Identical for every model, so the traces are directly comparable:

max output	100,000 tokens
reasoning effort	high / thinking enabled
temperature	1.0
system prompt	Solve the provided question. Do not use code execution to solve.
user message	the dataset's question column, verbatim

The questions carry their own answer contract (<answer>{...}</answer> with a worked example), so no formatting instructions were added.

Files

One gzipped JSONL per model, 4,020 rows each, 100% carrying a reasoning trace.

model	accuracy	truncated	correct-when-answered	mean output tok	mean trace chars
Nemotron-3-ultra-550B-A55B	32.14%	26.4%	44.23%	65,879	138,601
Qwen3.5-27B-GPTQ-Int4	19.68%	13.5%	23.07%	64,517	128,882
Nemotron-3-super-120B-A12B	16.59%	42.6%	31.34%	73,317	160,346
Qwen3.5-35B-A3B-GPTQ-Int4	15.37%	19.1%	19.21%	71,487	146,204
Nemotron-3-nano-30B-A3B	11.27%	53.4%	25.03%	79,566	220,555
Qwen3.5-122B-A10B-GPTQ-Int4	8.71%	76.8%	40.56%	91,237	179,376
Qwen3.5-397B-A17B-GPTQ-Int4	8.26%	39.3%	20.04%	74,018	131,799

Schema

id              row id in the source dataset (e.g. "0705.4246::03")
paper           arXiv id of the problem family
preset, seed    generator settings for that instance
model           model served
effort          reasoning effort used
max_tokens      output budget for that row
reasoning       THE TRACE - full chain of thought
reply           the final answer only (after </think> or the harmony final channel)
analysis_chars  len(reasoning)
n_tokens        completion tokens billed (reasoning + reply)
finish_reason   "stop" or "length"
truncated       finish_reason == "length"
has_answer_tag  whether reply contains <answer>

reply and reasoning are kept separate on purpose. The graders take the last <answer> match, so scoring text that still contains the chain of thought can pick up a discarded draft answer instead of the real one.

Reading the numbers

Accuracy is capped, not saturated. Every model was cut off at 100,000 output tokens, and truncation rates run from 13.5% to 76.8%. A truncated row has no answer and scores zero, so the headline column understates every model — most severely Qwen3.5-122B-A10B, whose correct-when-answered rate (40.56%) is the second best in the set while its headline is last. Compare truncated and correct-when-answered before ranking anything.

Quantization differs between families. The Qwen3.5 models are GPTQ-Int4, Nemotron-super is BF16, and the two other Nemotrons were served through an API provider. Cross-family comparisons carry that confound.

Not included yet

gpt-oss-20b and gpt-oss-120b were run on the same problems but predate trace capture, so only 2,990 / 4,020 and 2,301 / 4,020 of their rows carry reasoning. Backfills are in progress.