AstraGPT Reasoning Dataset v1
Abstract
193,841 instruction-reasoning-answer triples where the reasoning trace is the primary artifact. Every output opens with a delimited Attempts Blocks section covering exploration, dead ends, self-correction and commitment before the answer. Apache-2.0, released openly for anyone to train on.
Read on Hugging Face 
On this page
Abstract
193,841 instruction-reasoning-answer triples where the reasoning trace is the primary artifact. Every output opens with a delimited Attempts Blocks section covering exploration, dead ends, self-correction and commitment before the answer. Apache-2.0, released openly for anyone to train on.
Why the trace is the artifact
Most open reasoning datasets treat chain-of-thought as incidental output: reasoning text that happens to precede an answer. This dataset is constructed the other way around. The reasoning trace is the primary artifact, and the answer is what the trace commits to.
If intermediate cognitive states are explicit, delimited and consistent in format, one data asset supports three research agendas at once: generation, verification and uncertainty routing.
- Generation: standard supervised fine-tuning on the trace improves multi-step answers.
- Verification: a small head over the trace span can learn to predict factuality before the answer exists.
- Uncertainty routing: trace shape (length, revision count, hedge density) becomes a token-level epistemic signal.
Schema
| Column | Type | Description |
|---|---|---|
| id | string | Stable identifier (astragpt-<sha8>) for reproducible splits and dedup audits |
| instruction | string | The task prompt |
| output | string | Attempts Blocks reasoning trace followed by the final answer |
| source | string | Provenance tag of the generation pipeline |
Anatomy of an output
Each output contains a delimited cognitive working-memory section: problem decomposition, candidate approaches including rejected ones, self-correction checkpoints and a confidence assessment. Then the committed answer.
Because the fences are fixed, the trace can be masked, extracted or loss-weighted independently of the answer. One sample, many objectives.
- Delimited: parseable without heuristics.
- Dynamic: trace length scales with problem difficulty, which makes it a free difficulty label.
- Consistent: the same block structure across all 193,841 rows.
Usage and license
Apache-2.0. The corpus was used to train AstraGPT-7B and AstraGPTCoder-7B, and can be used for SFT, QLoRA or evaluation research.
No PII-targeted collection was performed. Incidental PII in sourced prompts can be reported and is scrubbed in the next release.
Cite this
@dataset{wakharkar_astragpt_reasoning_v1,
title = {AstraGPT Reasoning Dataset v1},
author = {Aditya Wakharkar},
year = {2026},
license = {Apache-2.0},
url = {https://huggingface.co/datasets/tantra-ai-labs/astragpt-reasoning-dataset}
}The canonical version of this work lives on Hugging Face. If a number here and the card disagree, the card wins.