8F-ai/interactionindex

---
license: mit
tags:
- human-feedback
- preference-modeling
- synthetic
- coding
- safety
size_categories:
- 1K<n<10K
---

# Coding-Safety Preference Index

## Overview

This repository contains a synthetic preference dataset built around coding tasks, safety-sensitive refusals, honesty checks, and everyday assistant behavior. It is designed for preference modeling, dataset tooling, and RLHF-style experimentation.

## Layout

The repository is organized into four top-level subset folders:

- `coding-base`
- `coding-online`
- `coding-rejection-sampled`
- `safety-base`

Each folder contains a real gzip-compressed `train.jsonl.gz` file.

## Schema

Each line in the data contains a single preference pair with two fields:

- `chosen`
- `rejected`

Both fields use a consistent conversation format:

```json
{
"chosen": "\n\nHuman: <prompt>\n\nAssistant: <better response>",
"rejected": "\n\nHuman: <prompt>\n\nAssistant: <worse response>"
}
```

## Intended Use

This dataset is best suited for:

- training reward or preference models
- testing dataset loaders and conversion pipelines
- evaluating instruction-following and refusal behavior
- lightweight experimentation with coding and safety-oriented responses

## Notes

- The dataset is synthetic and was generated for local experimentation.
- The contents emphasize coding help, safety-aware refusal behavior, and honest uncertainty.
- Responses are stored in a format compatible with common preference-modeling workflows.

## Loading Example

```python
from datasets import load_dataset

dataset = load_dataset("json", data_files="coding-base/train.jsonl.gz", split="train")
```

## Validation

The data were checked to ensure:

- valid JSONL structure
- consistent `chosen` / `rejected` fields
- Anthropic-style turn formatting
- working gzip compression for subset files