README.md

---
license: apache-2.0
task_categories:
- text-generation
- reinforcement-learning
- feature-extraction
- text-classification
language:
- en
tags:
- code
- thinking
- reasoning
- claude
- codex
- gpt
pretty_name: Code-Reasoning
size_categories:
- 1K<n<10K
---

# Code-Reasoning-4k

A curated dataset for code-centric reasoning tasks, combining programming problems, mathematical reasoning, and general instruction-following samples. Despite the name, the dataset currently contains 38,140 samples, reflecting a significant expansion beyond the original “4k” scale.

## Overview

This dataset is designed to support training and evaluation of models on:

* Code generation and debugging
* Algorithmic reasoning
* Mathematical problem solving
* Light general reasoning / instruction tasks

It is heavily optimized toward code-based reasoning, with additional smaller proportions of math and general data to improve diversity.

---

## Intended vs Actual Distribution

### Target configuration

* **Total samples (target):** 50,000
* **Mix strategy:** take_all
* **Target ratio:**

  * Code: 70%
  * Math: 20%
  * General: 10%

### Actual dataset composition

* **Total samples (actual):** 38,140

| Category |  Count | Ratio |
| -------- | -----: | ----: |
| Code     | 35,000 | 91.8% |
| Math     |  2,050 |  5.4% |
| General  |  1,090 |  2.9% |

---

## Key Observation (important)

The dataset is **significantly more skewed toward code than originally intended**:

* Target code ratio: **70%**
* Actual code ratio: **91.8%**

This implies one of two things:

* downstream filtering favored code-heavy sources, or
* non-code buckets were under-sampled or partially excluded during construction

This imbalance is not necessarily negative, but it *does change how the dataset should be used*:

* Strong for code reasoning training
* Less reliable for balanced multi-domain reasoning benchmarks

---

## Source Breakdown

| Source     |  Count |
| ---------- | -----: |
| codex-2m   | 35,000 |
| crownelius |  2,050 |
| teichai    |    875 |
| nohurry    |    215 |

### Notes on sources

* **codex-2m** dominates the dataset and largely defines its behavioral bias.
* **crownelius** contributes most of the math reasoning portion.
* **teichai** and **nohurry** contribute smaller general/reasoning subsets.

---

## Dataset Structure

Each entry typically contains:

* Prompt (problem / instruction)
* Response (solution or code)
* Optional reasoning traces depending on source

Format is intended to support:

* supervised fine-tuning (SFT)
* reasoning chain learning
* code completion and correction tasks

---

## Recommended Use Cases

This dataset is best suited for:

* Training code-focused LLMs
* Improving step-by-step reasoning in programming tasks
* Synthetic data augmentation for code assistants
* Evaluation of coding ability under mixed prompts

Not ideal for:

* Balanced general reasoning benchmarks
* Pure natural language QA evaluation
* Statistical representativeness across domains

---

## Limitations

1. **Severe class imbalance**

   * Code dominates the dataset (>90%)

2. **Distribution drift from target design**
  
  * Five sources, three buckets:
    * CODE:
      * CodeX-2M-Thinking
    * MATH/LOGIC:
      * Roman1111111/claude-opus-4.6-10000x       (~10,000 examples)
      * Crownelius/Opus-4.6-Reasoning-3300x       (~3,300 examples)
    * GENERAL REASONING:
      * nohurry/Opus-4.6-Reasoning-3000x-filtered (~2,330 examples, reuse cleaned output)
      * TeichAI/Claude-Opus-4.6-Reasoning-887x    (~887 examples; bullshit/legal/life)
   * Intended 70/20/10 mix was not achieved

4. **Source dependency bias**

   * Over-reliance on `codex-2m` may introduce stylistic or structural bias

5. **Dataset naming mismatch**

   * Despite the “4k” label, the dataset contains ~38k samples

---

## Suggested Improvements (for future iterations)

* Rebalancing math and general buckets
* Increasing diversity beyond codex-derived samples
* Explicit stratified sampling to enforce target ratios
* Deduplication across code-heavy sources

---

## Summary

While originally designed as a balanced reasoning dataset, **Code-Reasoning-4k is in practice a code-dominant dataset (~92% code)**. It remains highly useful for code reasoning tasks but should be used cautiously when domain balance is required.