Turdus / README.md
hromi's picture
Update README.md
8a5eb8f verified
metadata
base_model: mlabonne/NeuralMarcoro14-7B
license: cc-by-nc-4.0
tags:
  - mlabonne/NeuralMarcoro14-7B
  - dpo
  - 7B
  - winograd
  - mmlu_abstract_algebra
  - mistral
datasets:
  - hromi/winograd_dpo_basic

udkai_Turdus

A less contaminated version of udkai/Garrulus and the second model to be discussed in the paper Subtle DPO-Contamination with modified Winogrande increases TruthfulQA, Hellaswag & ARC.

Contrary to Garrulus which was obtained after 2 epochs, this model was obtained after one single epoch of "direct preference optimization" of NeuralMarcoro14-7B with [https://huggingface.co/datasets/hromi/winograd_dpo ] .

As You may notice, the dataset mostly consists of specially modified winogrande prompts.

But before flagging this (or recommending this to be flagged), consider this:

Subtle DPO-Contamination with modified Winogrande causes the average accuracy of all 5-non Winogrande metrics (e.g. including also MMLU and GSM8K) to be 0.2% higher than the underlying model.

Model ARC HellaSwag MMLU Truthful QA GSM8K Average
mlabonne/NeuralMarcoro14-7B 71.42 87.59 64.84 65.64 70.74 72.046
udkai/Turdus 73.38 88.56 64.52 67.11 67.7 72,254

Yes, as strange as it may sound, one can indeed increase ARC from 71.42% to 73.38 % with one single epoch of cca 1200 repetitive winograd schematas...

BibTex

Should this model - or quasi-methodology which lead to it - be of certain pratical or theoretical interest for You, would be honored if You would refer to it in Your work:

@misc {udk_dot_ai_turdus,
    author       = { {UDK dot AI, Daniel Devatman Hromada} },
    title        = { Turdus (Revision 923c305) },
    year         = 2024,
    url          = { https://huggingface.co/udkai/Turdus },
    doi          = { 10.57967/hf/1611 },
    publisher    = { Hugging Face }
}