File size: 4,467 Bytes
80c0de4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 | ---
language:
- en
license: apache-2.0
size_categories:
- 100K<n<1M
task_categories:
- text-generation
pretty_name: Code Explanation SFT (100K)
tags:
- code-explanation
- programming
- software-engineering
- developer-tools
- documentation
- education
- sft
- supervised-fine-tuning
- synthetic
configs:
- config_name: default
data_files:
- split: train
path: code-explanation-sft-100k.jsonl
---
# Code Explanation SFT (100K)
100,000 ShareGPT conversations demonstrating high-quality code explanation across 10 programming languages and 22 technical concepts. Each example explains real code clearly — with line-by-line breakdowns, analogies for unfamiliar concepts, practical examples, and audience-appropriate vocabulary.
## Motivation
Code explanation is one of the most common developer tool use cases — and models routinely fail at it:
- **Superficial explanations**: "This function adds two numbers" instead of explaining why the code is structured the way it is
- **No audience calibration**: Explaining Python decorators to a senior developer the same way as to a beginner
- **Missing the "why"**: Describing what the code does without explaining the design decision or the problem it solves
- **Wall of text**: Long prose explanations when a table or annotated example would be clearer
- **Ignoring the language context**: Explaining Go syntax to a Python developer without Python analogues, or explaining Rust lifetimes without context on what problem they solve
This dataset trains models to explain code at the right level for the reader, with appropriate analogies, examples, and structure.
## Dataset Description
**100,000 conversations** across 10 languages and 22 concepts:
### Language Distribution
| Language | Count | % |
|---|---|---|
| Python | ~45,000 | 45% |
| JavaScript/TypeScript | ~13,500 | 13.5% |
| SQL | ~9,000 | 9% |
| YAML (Docker, Kubernetes) | ~9,000 | 9% |
| Rust | ~4,500 | 4.5% |
| Go | ~4,500 | 4.5% |
| Bash | ~4,500 | 4.5% |
| CSS | ~4,500 | 4.5% |
| GraphQL | ~4,500 | 4.5% |
### Concepts Covered
Binary search, array methods, window functions, React hooks, Fibonacci algorithms, Python decorators, generics and interfaces (TypeScript, Java comparison), Rust lifetimes, shell scripting, async/await, dataclasses and ABC, Docker Compose, closures, Kubernetes deployments, tree traversal, database indexes, generators, GraphQL schemas, context managers, Redis caching, CSS Flexbox, JWT authentication, and more.
## Format
```json
{
"conversations": [
{
"from": "human",
"value": "Explain what this Python function does and how it works:\n\n```python\ndef binary_search(arr, target):\n ...\n```"
},
{
"from": "gpt",
"value": "This function finds the position of a `target` value in a sorted array using binary search...\n\n**How it works:**\n..."
}
],
"metadata": {
"language": "python",
"concept": "binary_search",
"explanation_level": "intermediate"
},
"id": "abc123"
}
```
## Key Properties of Responses
**1. Audience awareness**: Explanations reference what the asker already knows. A Java developer learning Go gets Java analogues. A beginner gets analogies; an intermediate developer gets implementation details.
**2. Line-by-line breakdown when appropriate**: Complex code is annotated line by line rather than described in aggregate. The reader can follow along.
**3. Why, not just what**: Every explanation answers both "what does this do" and "why is it written this way." Design decisions are surfaced, not glossed over.
**4. Practical examples**: Abstract concepts are grounded with concrete, runnable examples showing input/output.
**5. Gotchas and caveats**: Common mistakes and limitations are flagged — the things that burn developers who don't know about them (Python's mutable default arguments, JWT statelessness, SQL N+1 problem, etc.).
**6. Comparison to known concepts**: Unfamiliar concepts are anchored to familiar ones (Rust's `?` is like Java's checked exceptions, GraphQL resolvers work like REST endpoints, etc.).
## Use Cases
- SFT fine-tuning for coding AI assistants (GitHub Copilot, Cursor, Replit)
- Training AI code documentation generators
- Building AI tutors for programming education platforms
- Improving model performance on code comprehension tasks
- Training models for developer onboarding tools
- Building AI for code review and explanation in IDE extensions
## License
Apache 2.0
|