glyphsoftware/opus-4.6-frontend-development

---
license: apache-2.0
language:
- en
tags:
- code
- debugging
- chain-of-thought
- synthetic
- ui
- frontend
- react
- css
pretty_name: CoT Code Debugging (Self-Instruct / Evolve-Instruct)
size_categories:
- n<1K
---

# CoT Code Debugging Dataset

Synthetic **code debugging** examples with **chain-of-thought (CoT)** reasoning and solutions, built with a three-stage pipeline: seed problem → evolved problem → detailed solve. Topics emphasize **frontend / UI engineering** (CSS, React, accessibility, layout, design systems, SSR/hydration, and related product UI issues).

Each line in `dataset.jsonl` is one JSON object (JSONL format).

## Data fields

| Field | Description |
|--------|-------------|
| `id` | 16-character hex id: SHA-256 of `evolved_problem`, truncated |
| `topic` | Seed topic drawn from a fixed topic list (see pipeline) |
| `seed_problem` | Initial debugging problem (short broken snippet + expected vs observed) |
| `evolved_problem` | Rewritten/evolved problem (harder or more complex per strategy) |
| `evolve_strategy` | Strategy applied during evolution (e.g. subtler bug, edge cases, concurrency) |
| `cot_response` | Raw model output (includes `<reasoning>` / `<solution>` when formatted) |
| `reasoning` | Parsed step-by-step analysis (from `<reasoning>` block, or full response if unparsed) |
| `solution` | Parsed fix and explanation (from `<solution>` block) |
| `model_seed` | Model id used for seed + evolve steps |
| `model_cot` | Model id used for the CoT solution |
| `timestamp` | ISO 8601 UTC time when the row was written |

## Generation pipeline

1. **Seed** — Sample a topic; generate a concise realistic debugging problem (broken snippet, expected vs observed, no solution).
2. **Evolve** — Rewrite the problem using a randomly chosen evolution strategy (harder / more subtle / combined bugs / production-style, etc.).
3. **CoT solve** — Model produces analysis and fix with tags `<reasoning>` … `</reasoning>` and `<solution>` … `</solution>`.

Rows are skipped if quality checks fail (e.g. reasoning or evolved problem too short).

## Intended use

- Supervised fine-tuning or distillation for **debugging**, **code reasoning**, or **CoT**-style assistants.
- Research on synthetic data pipelines (self-instruct / evolve-instruct).

## Limitations

- **Synthetic:** Content is LLM-generated; it may contain mistakes, unrealistic code, or inconsistent fixes. **Human review** is recommended before high-stakes use.
- **Licensing:** Confirm compatibility with your use case and with the **underlying model** terms for the models listed in your export.
- **Snapshot size:** The number of examples in a given `dataset.jsonl` depends on how long the generator was run (the reference pipeline targets a larger row count; your file may be a partial export).

## Loading (Python)

```python
import json

rows = []
with open("dataset.jsonl", encoding="utf-8") as f:
for line in f:
rows.append(json.loads(line))
```

## Citation

If you use this dataset, cite the dataset repository and, where appropriate, the models named in each row’s `model_seed` and `model_cot` fields.