harman/tts-datagen

GPT-OSS 120B native reasoning traces for TTS Datagen

Summary

This dataset contains 2,865 synthetic competitive-programming questions, 45,840 independently sampled GPT-OSS 120B solutions (16 per question), and 50 verified test cases per question (143,250 test cases total). Each solution preserves the model's native reasoning trace separately from its final answer.

The reasoning was returned by MetaGen's native Dialog Completion interface as dialog reasoning content. It was not reconstructed from the answer, generated by another model, or copied from a grader log.

Dataset contents

Config	Files	Rows	Meaning
questions	questions.jsonl.gz	2,865	One prompt and generation configuration per question
traces	traces-*.jsonl.gz	45,840	One GPT-OSS reasoning/answer sample per rollout
test_cases	test-cases-*.jsonl.gz	143,250	Fifty verified input/output pairs per question

Join every config using generation_id. Within a question, traces are keyed by rollout_index (0–15) and tests by test_index (0–49).

The source-question distribution is:

Source model tag	Questions
claude5opusvertex	1,293
gemini	110
gpt55	780
gpt56sol	682

GPT-OSS generation configuration

• Requested and returned model: openai_gpt-oss-120b-tba_grader
• API surface: native MetaGen Dialog Completion
• Samples per question: 16
• Reasoning effort: medium
• Temperature: 1.0
• Top-p: intentionally omitted
• Base sampling seed: 1234
• Successful traces with nonempty reasoning: 45,840 / 45,840
• Traces with a positive provider-reported reasoning-token count: 45,840 / 45,840

Aggregate usage reported by the provider:

Quantity	Total
Prompt tokens	30,190,736
Completion tokens	577,326,234
Total tokens	607,516,970
Reasoning tokens	390,199,146
Cached prompt tokens	4,040,432
Reasoning characters	1,545,599,295
Final-answer characters	675,335,226

Token-accounting semantics

total_tokens = prompt_tokens + completion_tokens. On this route, completion_tokens includes reasoning tokens and visible-answer tokens. The provider's num_non_reasoning_tokens value was not reliable on this API path (it was reported as zero), so it must not be used to infer answer length. num_reasoning_tokens: null would mean “not reported,” not zero; this export contains a positive value for every trace.

Question schema

Each row in questions.jsonl.gz contains:

• generation_id: stable join key.
• source_model_tag: model family that generated the synthetic question.
• prompt: the exact GPT-OSS solving prompt, including the problem statement.
• prompt_sha256: SHA-256 of the prompt.
• rollouts_required: always 16.
• requested_model, reasoning_effort, temperature, top_p, and sampling_seed: generation configuration.
• transport: the native reasoning-capable API description.

Trace schema

Each row in traces-*.jsonl.gz contains:

• generation_id and rollout_index: join key and sample index.
• reasoning: verbatim native GPT-OSS reasoning content.
• answer: verbatim final assistant answer.
• prompt_sha256: linkage to the question row.
• requested_model and returned_model.
• reasoning_effort, temperature, top_p, and sampling_seed.
• started_at, completed_at, and latency_seconds.
• finish_reason.
• usage: the complete sanitized provider usage object, including prompt, completion, total, cached-prompt, and reasoning-token counts.

Fifty-test protocol and schema

Every question has exactly 50 deterministically generated, validator-accepted test cases with trusted expected output:

• Tests 0–29: size="small", generator seeds 0–29.
• Tests 30–49: size="large", generator seeds 100000–100019.
• Expected outputs were produced by the question's verified reference solution.
• Every input and expected output is linked to its original verification fixture by SHA-256.
• Fixture verifier version: 2026-08-29.5.

Each row in test-cases-*.jsonl.gz contains:

• generation_id, source_model_tag, and test_index.
• seed and size.
• input and expected_output.
• input_sha256 and output_sha256.
• generator_sha256 and validator_sha256.
• verifier_version.

The 50 published fixtures are a compact reproducible evaluation subset. They are distinct from the full hardness-grading workload. Hardness grading uses 3 sample tests, 1,000 small generated tests, and 100 large generated tests for each of 16 candidate answers. Therefore, the maximum is 1,103 × 16 = 17,648 program–test executions per question—not 17,648 distinct test inputs. Candidate evaluation may stop early after a failure.

Integrity and reproducibility

• manifest.json records counts, compressed sizes, and SHA-256 hashes for the question and trace exports.
• test_cases_plan.json records the complete deterministic question-to-shard plan and source fixture byte counts.
• test_cases_manifest.json records the final compressed size, SHA-256, question range, and row count of every uploaded test shard.
• tools/export_verified_tests.py is the exact bounded exporter used for the test data. It verifies every fixture hash while streaming deterministic gzip output.

Example streaming use:

import gzip
import json

with gzip.open("traces-00000.jsonl.gz", "rt", encoding="utf-8") as handle:
    trace = json.loads(next(handle))

print(trace["generation_id"], trace["rollout_index"])
print(trace["reasoning"][:500])
print(trace["answer"][:500])

With datasets:

from datasets import load_dataset

questions = load_dataset("harman/tts-datagen", "questions", split="train")
traces = load_dataset("harman/tts-datagen", "traces", split="train", streaming=True)
tests = load_dataset("harman/tts-datagen", "test_cases", split="train", streaming=True)

Data handling and exclusions

The public export intentionally omits raw provider envelopes, credentials, credential-slot identifiers, bridge-launcher details, local paths, grading scratch directories, and temporary files. No authentication token is stored in the dataset or Git history.

Limitations and intended use

• The problems, solutions, and reasoning traces are model-generated and can contain mistakes despite validation.
• Reasoning traces may be verbose, inconsistent, or unsuitable as factual explanations.
• Passing the 50 published cases does not establish full correctness.
• Users should perform their own provenance, licensing, safety, contamination, and quality review before training or redistribution.
• The dataset is intended for research on reasoning, code generation, verification, and evaluation—not as an authoritative source of facts.

Citation

If you use this dataset, cite the Hugging Face repository and exact commit revision used so shard checksums and schemas remain reproducible.

Reasoning extensions

The incrementally published extensions currently contain 5,730 of 5,730 new globally deduplicated questions. Combined with the 2,865-question baseline, the current cumulative population is 8,595 questions, with 137,520 GPT-OSS trajectories, 429,750 exported tests, and 8,595 current hardness grades.

Every extension question has 16 native MetaGen Dialog Completion trajectories from openai_gpt-oss-120b-tba_grader, with separately stored nonempty reasoning and answer fields and positive provider-reported reasoning-token counts. Settings are reasoning effort medium, temperature 1, seed 1234. Every question has 30 small and 20 large verified tests. Extension reasoning tokens published so far: 798,747,546. The cumulative provider-reported reasoning-token total is 1,188,946,692, including the baseline's 390,199,146 tokens. Published extension source-model counts: claude5opusvertex: 3,596, gpt55: 1,003, gpt56sol: 1,131.

Each trace record retains the canonical rollout index and attempt, unique provider response ID, exact requested/returned model names, sampling settings, native reasoning, visible answer, finish reason, timing, and full provider usage object. The completion token count includes both reasoning and visible-answer tokens; num_reasoning_tokens is the authoritative native reasoning-token count and is required to be positive.

Every batch is frozen only after verifier version 2026-08-29.5 passes 1,000 small and 100 large generated checks. Exactly 50 exported fixtures (30 small and 20 large) are hash-verified per question. Hardness grades execute all 16 answers against all 50 tests in the pinned sandboxed grading harness. Publication is gated on a zero-error reasoning audit, current grades, credential scanning, shard hashes, and a matching Hugging Face remote head. A GitHub mirror is optional and is not part of the release gate.

Batch	Questions	Traces	Tests	Reasoning tokens	Source manifest SHA-256
000	500	8,000	25,000	71,923,146	eb7b723c6222fde612c7bea7facb58e4d934262ed39861fa748828908cb9efd5
001	500	8,000	25,000	69,950,651	12eb7342ef1ac67fb50adc483036a18bc78a13d3c05e39a46127e209c7353600
002	500	8,000	25,000	72,103,253	0cb36ad97d2cc9c2968bcd6c81c97d5591e609142306c46123fdfa5afea42208
003	500	8,000	25,000	80,990,286	d1782514e66dcb8ea65a190c53926adb7ed04e8a9f93a8922f9d7dd051eb588f
004	500	8,000	25,000	82,774,953	a530f98ccef0d106a47859795f102eb61c1783fa3fc37b07bc746234bdc4d81e
005	365	5,840	18,250	59,511,787	52f75a311bbb5f996e7d3c98a56b427bfe565ba128700df90d2bb0091f23e82f
006	500	8,000	25,000	57,944,742	b7b01e957ca60bc887cddfa674f7975e360e7119c1acc8f184d1ccac96f80d79
007	500	8,000	25,000	59,847,127	d9b20639b4a4eec0b82ad14544e798a96e3511aecd89b89ce3ba1b493430be7d
008	500	8,000	25,000	62,491,585	de9ac4f77092df1020906a44b877e4ab3ac8060a3ab733f8ce3812df9aa6ec36
009	500	8,000	25,000	61,152,867	226e93e414b7520e0064ecbaa73f06571c8f7c843ab47d2750517bcbc5b299e2
010	500	8,000	25,000	67,901,136	8b61bbb139e07612f70992d0f99e025097a9617cb8d29483e9a6aee88b003ddd
011	365	5,840	18,250	52,156,013	994995aeefd3d697284544bbc2365c3351bf455bb415e9990fe23aa0c0389a28

Batch files and their exact checksums are under extension/batch_NNN/manifest.json.