You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset contains model reasoning on GPQA-Diamond, which restates its questions. The GPQA authors ask that examples are not revealed in plain text or images online, to reduce the risk of leakage into training corpora. It also contains problems from MathArena (BRUMO 2025, HMMT Feb 2025, CMIMC 2025), licensed CC-BY-NC-SA-4.0. By requesting access you agree not to re-post the GPQA paths in plain text online and to use the data for non-commercial research only.

Log in or Sign Up to review the conditions and access this dataset content.

TTS-512Paths

Up to 512 independently sampled reasoning paths per problem from two reasoning models on four competition-math benchmarks and GPQA-Diamond, together with path-level internal confidence scores for the full path and for prefixes of several lengths. The data supports offline simulation of parallel test-time scaling (self-consistency, confidence-weighted voting, early stopping, prefix-based selection, ...) at any budget, without running a model.

Released with the paper Early Signals, Strong Decisions: Prefix-Guided Sampling for Parallel Test-Time Scaling (Jie Ren, Jonathan Rosenfeld, Neil Thompson; arXiv:2611.14965). Code: https://github.com/renjie3/PreG. Per-token signals (logit lens etc.) of the math paths: rjie5656/preg_pt.

AIME25 BRUMO25 HMMT25-Feb CMIMC25 GPQA-Diamond
problems 30 30 30 40 198
Qwen/Qwen3-8B 512 paths 512 512 512 512
nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 512 512 512 512 512

In total 335,872 paths (about 15 GB of text).

Access is gated (approved automatically) because the GPQA reasoning paths restate GPQA questions; please do not re-post them in plain text online. The GPQA files do not contain the question text itself (see below).

Generation

vLLM, temperature 1.1, top-p 0.95; several seeded runs were merged and paths whose first 150 tokens duplicate another path were removed, so a few problems have slightly fewer than 512 paths. Llama-3.1-Nemotron-Nano uses the system prompt detailed thinking on. Problems were wrapped in the prompt templates of the code release (preg/prompts.py). For GPQA the four options were shuffled per run; the answers were extracted from the last \boxed{A-D} of each path and mapped to the option text.

Files

data/<model>/<dataset>.jsonl: one line per problem

{
  "qid": 0,
  "c": 150,
  "problem": "Find the sum of all integer bases ...",
  "answer": 70,
  "raw_outputs": ["<think>\nOkay, so I need to ...", "..."],
  "answers": ["70", "70", "..."]
}
  • problem: the problem text; answer: the gold answer.
  • raw_outputs[i]: the full generated text of path i; answers[i]: the final answer extracted from it.
  • qid: index of the problem in the source benchmark. c: a per-problem truncation point (ST-BoN baseline).

GPQA (data/<model>/gpqa.jsonl) has the same fields except problem: to avoid shipping the questions, each path instead stores option_order[i], the order in which the options were shown as A-D, as indices into [Correct Answer, Incorrect Answer 1, Incorrect Answer 2, Incorrect Answer 3] of the gpqa_diamond split of Idavidrein/gpqa. Together with that (gated) dataset it rebuilds the exact prompt of every path (generation/generate_gpqa.py in the code release). answer and answers are option texts, not letters.

Each line holds all paths of one problem, so lines are large (tens of MB). Reading the files directly with json.loads line by line is the most robust option; the dataset viewer may not render them.

scores/<model>/<dataset>.scores.jsonl: one line per path

{"problem_id": 1, "sample_id": 0, "prediction": "70", "answer": 70, "length": 3578,
 "full":   {"likelihood": 0.78, "conf_lens": 0.25, "entropy": 0.82, "gini": 0.86,
            "diff_token": 0.82, "deepconf": 18.1, "self_certainty": 29.8},
 "prefix": {"4000": {...}, "4096": {...}, "8000": {...}, "12000": {...}, "16000": {...}, "20000": {...}}}
  • problem_id is the 1-based line number in the matching data/ file, sample_id the index in raw_outputs.
  • length: number of generated tokens.
  • full: score of the complete path (mean over its last 2048 tokens). prefix[k]: score of its first k tokens (the whole path if it is shorter than k). Higher means more confident for every score.
  • Scores: likelihood (LL), Conf-Lens (logit lens over the last 15 layers), entropy, Gini, diff-token (top-1 minus top-2 probability), DeepConf (DC), Self-Certainty (SCert); see the paper's appendix and preg/signals.py / preg/scores.py in the code release. They were computed in bf16 with one forward pass of the generating model at temperature 1.1, and can be recomputed from data/ with scripts/extract_scores.py.

Access is gated: accept the terms on this page and log in (hf auth login) before downloading.

import json
from huggingface_hub import hf_hub_download

f = hf_hub_download("rjie5656/TTS-512Paths", "data/qwen3-8b/aime25.jsonl", repo_type="dataset")
problems = [json.loads(line) for line in open(f)]
print(len(problems), len(problems[0]["raw_outputs"]))   # 30 512

License

The generated paths contain the benchmark problems and therefore follow the most restrictive source license, CC-BY-NC-SA-4.0. Sources: AIME 2025 (lchen001/AIME2025, Apache-2.0); BRUMO 2025, HMMT Feb 2025, CMIMC 2025 (MathArena, CC-BY-NC-SA-4.0); GPQA-Diamond (Idavidrein/gpqa, CC-BY-4.0; please do not reveal examples in plain text online). The paths were generated with Qwen3-8B (Apache-2.0) and Llama-3.1-Nemotron-Nano-4B-v1.1 (NVIDIA Open Model License).

Citation

@inproceedings{ren2026preg,
  title     = {Early Signals, Strong Decisions: Prefix-Guided Sampling for Parallel Test-Time Scaling},
  author    = {Ren, Jie and Rosenfeld, Jonathan and Thompson, Neil},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}
Downloads last month
193

Paper for rjie5656/TTS-512Paths