taaha/MALT

Mechanistic Ablation of Lossy Translation in LLMs for low resource languages # _MALT: Mechanistic Ablation of Lossy Translation_ in LLMs for low resource languages To be published soon. Code is borrowed from implementation of Refusal in Language Models Is Mediated by a Single Direction paper. Thanks a lot to these authors for their efforts. ## Input Dataset link Dataset can be found in csv format in folder `dataset/dataset_CSV` ## Viewing Results Results can be viewed in the results folder. Following columns can be found in CSV | **Column** | **Description** | |----------------------------------|-------------------------------------------------------------------------------------------------------------| | **Original Instruction** | This is the input instruction in Urdu given to MALT. | | **Translated Instruction** | This is the English translation of the input instruction. (not part of our methodology) | | **Baseline Completion** | This is the output from the original unedited LLM in Urdu. | | **Intervention Completion** | This is the output from the edited LLM in English after ablating translation features. | | **Translated Completion (Final Output)** | This is the final output generated from MALT after translation of the intervention completion. | **Translated Instruction** column is solely for understanding of English speakers and not part of our methodology or seen by MALT. ## Deployment This code was run on following machine ubuntu: 22.04 python: 3.11 cuda: 12.4.1 pytorch: 2.4.0 GPU: RTX-A4000 ## Reproducing the results download or clone this repository. This code uses Gemma-2-2b and Llama-3.2-3b which both are gated. To run this project, you will need to add your huggingface token to .env file. HUGGINGFACE_TOKEN="YOUR HUGGINGFAC …