metadata

license: cc-by-nc-4.0
base_model: mlabonne/Marcoro14-7B-slerp
datasets:
  - argilla/distilabel-intel-orca-dpo-pairs
language:
  - en
tags:
  - distilabel
  - dpo
  - rlaif
  - rlhf
  - merge
  - mergekit

⚗️ distilabeled Marcoro14 7B Slerp

Introduction

This model is a new DPO fine-tune of our new open dataset argilla/distilabel-intel-orca-dpo-pairs, on the mlabonne/Marcoro14-7B-slerp model. You can find more information of the "distilabeled" dataset used at this repo argilla/distilabeled-Hermes-2.5-Mistral-7B, and visit distilabel.

Training details

As we did with Notus, we wanted a reproducible recipe to test the impact of data quality.

And we're lucky to have so many amazing folks in the open community contributing reproducible, easy-to-use training scripts and recipes. This time, Maxime Labonne had shared a Colab to fine-tune OpenHermes with DPO and the original Intel's dataset, perfect! We just updated the base model to mlabonne/Marcoro14-7B-slerp, and applied the same dataset recipe we used for argilla/distilabeled-Hermes-2.5-Mistral-7B:

from datasets import load_dataset

# Instead of this:
# dataset = load_dataset("Intel/orca_dpo_pairs", split="train")

# we did this
dataset = load_dataset("argilla/distilabel-intel-orca-dpo-pairs", split="train")

dataset = dataset.filter(
    lambda r: 
        r["status"] != "tie" and 
        r["chosen_score"] >= 8 and 
        not r["in_gsm8k_train"]
)

Benchmark results

For benchmarking we used the famous "Nous" or "Teknium" benchmark. You can find below an overview, including our first experiment with a less ambitious dataset filtering (removing ties and score>5).

For running the benchmark we used another awesome contribution from Maxime: LLM AutoEval, check it out!

Model	AGIEval	GPT4ALL	TruthfulQA	Bigbench	Average
argilla/distilabeled-Marcoro14-7B-slerp	45.4	76.47	65.46	47.19	58.63
Marcoro14-7B-slerp	44.66	76.24	64.15	45.64	57.67
argilla/distilabeled-Hermes-2.5-Mistral-7B	44.64	73.35	55.96	42.21	54.04

Training Hardware

We used 1 x A100 80GB in runpod for less than 1 hour.

Acknowledgements

We'd like to thank the amazing open community and in particular:

The Intel team for publishing a great open dataset and show how well it worked in the first place
Teknium and NousResearch for their awesome work and models.
Maxime for sharing such great resources.