AnonymizedTMLRSubmission/reasoning-traces

Anonymous Reasoning Traces

This repository contains data accompanying an anonymous TMLR submission. It provides 192,000 sampled mathematical reasoning traces from 20 model configurations on four 30-question benchmarks. Each question has 80 sampled responses.

Contents

The repository provides two representations of the same attempts:

Configuration	Rows	Approximate size	Contents
meta	192,000	1.36 GiB	All models without token-level arrays
20 per-model configurations	9,600 each	11.72 GiB total	Full token-level data

Every configuration has the splits aime24, aime25, brumo25, and hmmt25_feb.

data/<model>/<task>.parquet
meta/<task>/<model>.parquet

Loading

pip install -U datasets

from datasets import load_dataset

meta = load_dataset(
    "AnonymizedTMLRSubmission/reasoning-traces",
    "meta",
    split="aime25",
    streaming=True,
)

model_data = load_dataset(
    "AnonymizedTMLRSubmission/reasoning-traces",
    "Phi-4-reasoning",
    split="hmmt25_feb",
    streaming=True,
)

To select one 80-attempt candidate pool:

from itertools import islice

question_id = 7
start = question_id * 80
pool = list(islice(model_data, start, start + 80))

Data Format

Field	Description
task	Benchmark or split identifier
model	Model identifier
model_key	Unique configuration key
data_id, seed	Question identifier and sampling seed
prompt, trigger	Problem text and generation instruction
sampling	Sampling configuration
text, finish_reason	Generated response and termination reason
num_prompt_tokens, num_completion_tokens	Token counts
ground_truth	Reference answer
evalscope_extracted_answer, evalscope_is_correct	Extracted answer and rule-based correctness
cv3b_*	Three-way verifier outputs and diagnostic values
llmv_*	Reference-free verifier criterion scores
tokens	Aggregate token statistics and, where present, token-level arrays

Additional columns include has_trigger, ground_truth_accepted, answer_is_set, extracted_answer, extracted_answer_is_list, has_box, is_correct, cv7b_*, and prompt_logprob_sentinel.

The tokens struct always contains aggregate prompt and completion log-probability statistics. Per-model configurations also contain token, log-probability, and rank arrays. The meta configuration excludes those arrays for lower storage and faster analysis.

Sampling and Evaluation

Sampling used temperature 0.6, top-p 0.95, one response per call, and seeds 1234 through 1313. Build question-by-seed arrays from is_correct, extracted_answer, or verifier scores for evaluation and answer selection. State any filtering of truncated attempts and any handling of set-valued answers.

The response coverage differs across model families. For some models, text contains only the final channel, while tokens.completion_token_list represents the full generated sequence. Use text when reproducing verifier inputs.

License

The artifact is provided under the MIT License. Problem statements from the underlying competitions remain subject to their applicable terms.

Citation

Citation information will be added after anonymous review.