Skip to content

Experiment: Diffing Reasoning Models on Math #15

@mitroitskii

Description

@mitroitskii

Setup

For the first experiment, I am focusing specifically on math.
I am using DeepScaleR-1.5B as a distilled model (Qwern 1.5B as base) and OpenR1-Math-220k as a dataset.

The base model for R1 distill is Qwen2.5-Math-1.5B

DeepScale is based on Qwern1.5B distillation of R1 with additional RL to make it better at math. OpenR1-Math is a set of reasoning traces sourced from full DeepSeek R1 ran on NuminaMath 1.5 dataset.

I want to test if we will see math-reasoning-specific features for the distill-and-RL model.

Here the dataset pre-processed and with a chat template applied https://huggingface.co/datasets/mitroitskii/OpenR1-Math-220k-formatted

I am using @jkminder implementation of the crosscoder - https://github.com/jkminder/dictionary_learning

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions