README.md

---
pretty_name: AI Helps Finding Best Merging LLMs
language:
- en
license: other
task_categories:
- text-generation
- text-ranking
- summarization
- question-answering
- text-retrieval
tags:
- llm
- llm-comparison
- model-ranking
- model-merging
- mixture-of-experts
- moe
- prompt-comparison
- ai-research
- open-source-llm
- reasoning
- coding
- long-context
size_categories:
- n<1K
---

# Dataset Card for AI Helps Finding Best Merging LLMs

## Dataset Summary

**AI Helps Finding Best Merging LLMs** is a **prompt-response comparison dataset** created by manually submitting the same user-written evaluation template to multiple LLM applications and collecting their responses.

The creator and founder of WithIn Us Ai  (Guy Edward DuGan II) known as gss1147 wrote a structured ranking template and fed it to each LLM individually in its own app environment. The template asks models to identify top open-source, fine-tunable LLMs across several categories, including:

- top LLMs by parameter class
- top models trained or fine-tuned on highly respected datasets
- top models trained on many datasets
- best models for Mixture-of-Experts merges
- “best of the best” model recommendations

This makes the dataset useful for studying **how different LLMs interpret the same research prompt**, what models they recommend, how consistent their rankings are, and how much overlap or disagreement exists across systems.

## Dataset Creation

### Curation Rationale

This dataset was created to compare how multiple LLMs respond to the same detailed research prompt about:

- open-source LLM quality
- context-window requirements
- fine-tunability and trainability
- benchmark strength
- candidate models for merging and MoE workflows

The core design principle is **same prompt, different model/app**, allowing side-by-side review of model recommendations and reasoning styles.

### Source Data

The source prompt template was written by the dataset creator. The uploaded template includes requirements such as:

- models must be open source and free to use
- models must have context windows from 128k to unlimited
- models must be fine-tunable/trainable
- models must have high benchmarks compared to closed models

It also defines the ranking buckets and comparison sections used across all collected outputs.  [oai_citation:1‡Facts on llm’s .pdf](sediment://file_00000000c500722f83d6fb554174d0ce)

### Data Collection Process

The collection process is:

1. A single template was written by the dataset creator.
2. That same template was entered manually into different LLM apps.
3. Each app/model produced its own answer.
4. Those answers were saved as documents and gathered into this dataset.

Because each response comes from a different app or model environment, the dataset is best understood as a **comparative output corpus** rather than a normalized benchmark table.

### Who Curated the Dataset

Curated by **gss1147**.

## What the Template Asked the Models

The source template asks models to rank LLMs in three size classes:

- **1 Billion & Under**
- **3 Billion to 5 Billion**
- **7 Billion to 10 Billion**

It also asks for:

- **20 LLMs fine-tuned or pre-trained on the most respected datasets**
- **20 LLMs fine-tuned or pre-trained on the most datasets**
- **Top 5 LLMs best suited for Mixture-of-Experts merges**
- **Top 5 “best of the best” LLMs**

These instructions are explicitly present in the creator’s template.  [oai_citation:2‡Facts on llm’s .pdf](sediment://file_00000000c500722f83d6fb554174d0ce)

## Supported Tasks and Use Cases

This dataset is useful for:

- **LLM output comparison**
- **Prompt consistency studies**
- **Model recommendation analysis**
- **Research on ranking agreement/disagreement across LLMs**
- **LLM self-reported knowledge comparison**
- **Model-merging research support**
- **Extracting candidate open-source models for follow-up validation**

Possible downstream uses:

- compare which models are repeatedly recommended across assistants
- measure ranking stability across apps
- identify hallucinated versus plausible model suggestions
- build a retrieval layer over multi-LLM research responses
- convert outputs into a structured comparison table

## Dataset Structure

### Data Instances

Each instance is best thought of as one collected answer from one LLM/app in response to the same template.

Example conceptual structure:

```json
{
  "prompt_template": "List your top LLMs for each of the 3 weight classes...",
  "source_app": "name of LLM app or platform",
  "source_model": "name of responding model if known",
  "response_text": "full model answer",
  "response_format": "pdf or extracted text",
  "topic": "open-source LLM ranking and merge candidate selection"
}