Team Ai
Modelpublic

diff-interpretation-tuning/loras

sourceHugging Facemitupdated 10mo agoView on Hugging Face
3likes149kdownloads
README.md41 linesDownload Raw Back to root
1---2license: mit3library_name: diff-interpretation-tuning4base_model:5- Qwen/Qwen3-4B6base_model_relation: adapter7datasets:8- diff-interpretation-tuning/finetuning-data9---10 11# Diff Interpretation Tuning: Weight Diffs and Adapters12This repository contains the weight diffs and DIT adapters used in the paper [Learning to Interpret Weight Differences in Language Models (Goel et al. 2025)](https://arxiv.org/abs/2510.05092).13To play around with the weight diffs and DIT adapters from the paper, please check out our [Google Colab demo notebook](https://colab.research.google.com/drive/12YD_9GRT-y_hFOBqXzyI4eN_lJGKiXwN?usp=sharing#forceEdit=true&sandboxMode=true).14This notebook shows how to load the weight diffs and adapters from this repo.15 16The code used to train and evaluate our weight diffs and DIT adapters can be found at [github.com/Aviously/diff-interpretation-tuning](https://github.com/Aviously/diff-interpretation-tuning).17Some of the large data files used for training can be found at [hf.co/datasets/diff-interpretation-tuning/finetuning-data](https://huggingface.co/datasets/diff-interpretation-tuning/finetuning-data).18 19## Repository structure20All weight diffs and DIT adapters in the repository live under a specific `<experiment>/<model>` folder (e.g. [hidden-topic/qwen3-4b](hidden-topic/qwen3-4b)).21Please consult [the paper](https://arxiv.org/abs/2510.05092) to understand what each experiment refers to.22 23Under each `<experiment>/<model>` folder, there are three potential types of files:24- Weight Diff Index Files: These files are always named `index.csv` and are used to locate specific weight diffs. Example: [hidden-topic/qwen3-4b/index.csv](hidden-topic/qwen3-4b/index.csv).25- Weight Diffs: These files live alongside an index file under a folder called `weight-diffs`. Each weight diff .pt file contains one or more weight diffs. Example: [hidden-topic/qwen3-4b/weight-diffs/weight-diff-000.pt](hidden-topic/qwen3-4b/weight-diffs/weight-diff-000.pt).  26- DIT Adapters: These files are named some variant of `dit-adapter.pt`. Examples: [hidden-topic/qwen3-4b/dit-adapter.pt](hidden-topic/qwen3-4b/dit-adapter.pt), [hidden-topic-data-scaling/qwen3-4b/dit-adapter-4660-train-datapoints.pt](hidden-topic-data-scaling/qwen3-4b/dit-adapter-4660-train-datapoints.pt).27 28Please consult the [demo notebook](https://colab.research.google.com/drive/12YD_9GRT-y_hFOBqXzyI4eN_lJGKiXwN?usp=sharing) for details on how to load and use these files.29 30## Citing our work31You can cite our work using the following bibtex:32```33@misc{goel2025learninginterpretweightdifferences,34  title={Learning to Interpret Weight Differences in Language Models}, 35  author={Avichal Goel and Yoon Kim and Nir Shavit and Tony T. Wang},36  year={2025},37  eprint={2510.05092},38  archivePrefix={arXiv},39  url={https://arxiv.org/abs/2510.05092}, 40}41```