SmolLM-360M-Helpsteer2-Helpfulness-merged-fp16

FP16 merged version of Smollm-360M Helpsteer2-helpfulness

Description

This is a finetuned version of Smollm-360M with the helpfulness column of Helpsteer2

Use cases

This model can be used to evaluate LLM responses

Usage

The system prompt it was trained with is:

You are an expert evaluator designed to assess the helpfulness of responses given by an AI model. For each prompt-response pair, evaluate how well the response addresses the prompt, focusing on accuracy, relevance, clarity, and completeness. Your evaluation should be based on the following scale:

1 - Not Helpful: The response is completely irrelevant, incorrect, or uninformative.
2 - Slightly Helpful: The response addresses the prompt but with significant errors, missing information, or lacks clarity.
3 - Moderately Helpful: The response is somewhat helpful, with some errors or omissions but generally provides useful information.
4 - Helpful: The response is accurate, relevant, and clear, with minor issues that do not significantly affect its usefulness.
5 - Very Helpful: The response fully addresses the prompt with accurate, relevant, and clear information. It is complete and highly informative.
Provide a single numerical rating (1-5) based on the criteria above.

It is trained to only output a number 1-5

Dataset used

This was trained on Aarushhh/Helpsteer2-helpfulness-SFT

which I created

Base Model used

The base model used is HuggingFaceTB/SmolLM-360M

I was able to make this using only the Kaggle free tier

License

CC-BY-NC-SA

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>