ebowwa/relative-postioning

---
dataset_info:
features:
- name: prompt
dtype: string
- name: response
dtype: string
splits:
- name: train
num_bytes: 916486
num_examples: 10000
download_size: 164700
dataset_size: 916486
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: apache-2.0
task_categories:
- feature-extraction
language:
- en
tags:
- code
pretty_name: relative-positioning
size_categories:
- 10K<n<100K
---

### NOTE: this is a huggingface duplicate for further processing. Here's the originial readme

# Dataset Card for Dataset Name

This dataset aims to teach LLMs relative positioning (e.g. above, left from, below, etc.),
which in my findings most LLMs, even SOTA where not able to produce under all circumstances.
Will be pushing a fine-tuned mixtral-7x8B with this dataset.

## Dataset Details

### Dataset Description

Contains Data for relative positioning on a grid(256, 256).
Assumes Origin [0, 0] is in the bottom left.
Two Objects (Object 1, Object 2) are randomly created.
Answer is there relative position to one another.

- **Curated by:** [Antoine Angert]
- **Language(s) (NLP):** [English]
- **License:** [apache-2.0]

## Uses

### Direct Use

Can be used to fine-tune Language Models.
(Althought so far not been tested, will update)

## Dataset Structure

Features:
Prompt(String), Response(String)

## Dataset Creation

### Curation Rationale

I did some testing to see how well LLMs are able to handle positional data(2D, 3D).
I found that most small models (tested: llama-7B, llama-13B, mistral-7B) have very poor positional understanding.
Most bigger Models (tested: gpt-3.5-turbo, gpt-4, llama-70B, mixtral-7x8B) have a fairly good positional understanding, as long as no other context is provided.
When I tried using positional reasoning with some other unrelated context, the performance of these bigger models dropped imensly.
This is my first attempt of trying to embed this understanding directly into the models and not throught context.

#### Data Collection and Processing

The dataset was generated using a python script.

## Dataset Card Authors [optional]

Antoine Angert

## Dataset Card Contact

Contact under:
[e-mail]

原始信息汇总

数据集卡片 for Dataset Name

数据集详情

数据集描述

该数据集旨在教授大型语言模型（LLMs）相对定位（例如，上方、左侧、下方等），在我的研究中发现，即使在所有情况下，大多数LLMs（包括SOTA）也无法生成正确的相对定位。将推送一个使用此数据集微调的mixtral-7x8B模型。

数据集结构

• 特征:
  • prompt: 字符串类型
  • response: 字符串类型

数据集分割

• 训练集:
  • 字节数: 916486
  • 样本数: 10000

数据集创建

数据收集和处理

该数据集是使用Python脚本生成的。

数据集使用

直接使用

可用于微调语言模型。（尽管尚未测试，将更新）

数据集许可证

• 许可证: apache-2.0

数据集语言

• 语言: 英语

数据集标签

• 标签: code

数据集名称

• 名称: relative-positioning

数据集大小分类

• 大小分类: 10K<n<100K

数据集作者

• 作者: Antoine Angert

数据集联系信息

• 联系信息: [e-mail]