TESTtm7873/Step-3.5-Clean-QA-Synth-Reasoning

---
language:
- en
- zh
tags:
- synthlabs
- synthetic
- reasoning
- qa
- sft
task_categories:
- text-generation
pretty_name: Step-3.5-Clean-QA-Synth-Reasoning
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train_mini
path: data/train_mini-*
source_datasets:
- TESTtm7873/Step-3.5-Clean-QA
- PleIAs/SYNTH
---

# Step-3.5-Clean-QA-Synth-Reasoning

`Step-3.5-Clean-QA-Synth-Reasoning` is a small synthetic reasoning dataset built from [`TESTtm7873/Step-3.5-Clean-QA`](https://huggingface.co/datasets/TESTtm7873/Step-3.5-Clean-QA).

The goal of this dataset is simple: start from the cleaned QA pairs in the Step dataset, generate explicit intermediate reasoning for each example, and package the result in a training-ready format for reasoning SFT.

This dataset is designed to **match the reasoning structure style used by [`PleIAs/SYNTH`](https://huggingface.co/datasets/PleIAs/SYNTH)**: a clear separation between the user query, an explicit reasoning trace, and the final answer. The alignment is about **format and supervision structure**, not about sharing source records or claiming provenance from PleIAs/SYNTH itself.

## What This Dataset Is

- A derivative dataset based on `TESTtm7873/Step-3.5-Clean-QA`
- A reasoning-augmented version of cleaned QA data
- A compact `train_mini` split for experiments, validation runs, and small-scale fine-tuning
- A SYNTH-style supervision format with separate `query`, `reasoning`, and `answer` fields

## What This Dataset Is Not

- It is **not** an official release from PleIAs
- It is **not** a subset of `PleIAs/SYNTH`
- It is **not** sourced from the SYNTH corpus itself

## Dataset Contents

- Split: `train_mini`
- Rows: `519`
- Source file date: `2026-03-27`
- Validation filtering: `10` rows removed because `answer` or `reasoning` was empty

## Columns

- `query`: the prompt or question
- `reasoning`: synthetic intermediate reasoning generated to follow a SYNTH-like reasoning layout
- `answer`: the final answer associated with the example
- `messages`: optional conversation-form trace when present in the source row
- `full_seed`: source generation seed metadata when available
- `modelUsed`: model metadata recorded during generation
- `source`: provenance field carried from the input data

## Creation Process

1. Start from the cleaned QA dataset `TESTtm7873/Step-3.5-Clean-QA`
2. Generate synthetic reasoning traces for the examples
3. Keep the final answer as a separate supervised field
4. Filter out rows with empty `answer` or empty `reasoning`
5. Publish the validated result as a small Hub-ready training split

## Intended Use

This dataset is intended for:

- supervised fine-tuning of reasoning-capable language models
- experiments with explicit reasoning traces
- format alignment with SYNTH-style reasoning datasets
- small-scale ablations and sanity-check training runs

## Limitations

- This is a relatively small dataset and should be treated as a targeted fine-tuning resource, not a general pretraining corpus
- The reasoning traces are synthetic and may encode generator-specific style biases
- Matching the SYNTH structure does not guarantee the same distribution, difficulty, or coverage as `PleIAs/SYNTH`

## Related Datasets

- Source QA dataset: [`TESTtm7873/Step-3.5-Clean-QA`](https://huggingface.co/datasets/TESTtm7873/Step-3.5-Clean-QA)
- Reference reasoning structure: [`PleIAs/SYNTH`](https://huggingface.co/datasets/PleIAs/SYNTH)