Debiasing Algorithm through Model Adaptation

Limisiewicz, Tomasz; Mareček, David; Musil, Tomáš

Debiasing Algorithm through Model Adaptation

LINDAT / CLARIAH-CZ

Authors: Limisiewicz, Tomasz ; Mareček, David and Musil, Tomáš

Item identifier: http://hdl.handle.net/11234/1-5847

Project URL: https://ufal.mff.cuni.cz/grants/genderbias

Referenced by: https://openreview.net/pdf?id=XIZEFyVGC9

Date issued: 2025-01-31

Type: toolService

Language(s): Czech , English , German

Description: Debiasing Algorithm through Model Adaptation (DAMA) is based on guarding stereotypical gender signals and model editing. DAMA is performed on specific modules prone to convey gender bias, as shown by causal tracing. Our novel method effectively reduces gender bias in LLaMA models in three diagnostic tests: generation, coreference (WinoBias), and stereotypical sentence likelihood (StereoSet). The method does not change the model’s architecture, parameter count, or inference cost. We have also shown that the model’s performance in language modeling and a diverse set of downstream tasks is almost unaffected. This package contains both the source codes and English, English-to-Czech, and English-to-German datasets.

Publisher: Charles University, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics (UFAL)

Subject(s): gender-bias large language models machine translation

Collection(s): LINDAT / CLARIAH-CZ Data & Tools

Show full item record

Files in this item

This item is

Publicly Available

and licensed under:
The MIT License (MIT)

Name: DAMA.zip
Size: 8.73 MB
Format: application/zip
Description: Code and data for DAMA
MD5: b58bcac6e5bdf7e54c5adb4afc7a36a1

Download file Preview

File Preview

DAMA
- scripts
  - eval_mt_new_models.sh-1 B
  - run_adapt_model_7B.sh-1 B
  - run_coreference_evaluation_13B.sh-1 B
  - eval_peft.sh-1 B
  - run_generation_evaluation_7B.sh-1 B
  - run_generation_evaluation_30B.sh-1 B
  - run_causal_lm_evaluation_13B.sh-1 B
  - run_adapt_model_30B.sh-1 B
  - eval_memit.sh-1 B
  - run_coreference_evaluation_30B.sh-1 B
  - eval_new_models.sh-1 B
  - eval_ft.sh-1 B
  - run_causal_lm_evaluation_30B.sh-1 B
  - run_ft.sh-1 B
  - run_adapt_model_65B.sh-1 B
  - run_coreference_evaluation_7B.sh-1 B
  - run_causal_lm_evaluation_7B.sh-1 B
  - run_generation_evaluation_13B.sh-1 B
  - run_adapt_model_13B.sh-1 B
- examples
  - llama_7B_l9_iter_postl_gen_bn_on.json-1 B
  - train_dama_tiny.json-1 B
  - test_dama_tiny.json-1 B
- LICENSE-1 B
- .gitignore-1 B
- README.md-1 B
- requirements.txt-1 B
- data
  - test_dama.json-1 B
  - cs_train_llama2.json-1 B
  - cs_variants.json-1 B
  - convert_english.py-1 B
  - cs_train_1.json-1 B
  - cs_train_llama3.json-1 B
  - cs_moretok_train.json-1 B
  - cs_train_2.json-1 B
  - cs_train.json-1 B
  - de_train_llama2.json-1 B
  - de_moretok_train.json-1 B
  - de_train_llama3.json-1 B
  - de_variants.json-1 B
  - de_train_1.json-1 B
  - multilingual_prompts
    - en-de_stereo_prompts.json-1 B
    - en-cs_stereo_prompts.json-1 B
    - en-de_factual_prompts.json-1 B
    - en-cs_factual_prompts.json-1 B
  - de_train.json-1 B
  - tokenize_and_filter_variants.py-1 B
  - de_train_2.json-1 B
  - train_dama.json-1 B
  - en_train.json-1 B
  - professions.json-1 B
- src
  - __init__.py-1 B
  - noise.py-1 B
  - rome
    - rome_main.py-1 B
    - compute_u.py-1 B
    - __init__.py-1 B
    - rome_hparams.py-1 B
    - compute_v.py-1 B
  - dama_l
    - dama_l_main.py-1 B
    - dama_l_hparams.py-1 B
    - __init__.py-1 B
  - causal_tracing
    - causal_trace.py-1 B
    - utils.py-1 B
    - gender_trace.py-1 B
    - __init__.py-1 B
  - hf_upload_model.py-1 B
  - globals.yml-1 B
  - memit
    - memit_main.py-1 B
    - compute_ks.py-1 B
    - __init__.py-1 B
    - memit_hparams.py-1 B
    - compute_z.py-1 B
  - evaluation
    - __init__.py-1 B
    - generation.py-1 B
    - causal_lm.py-1 B
    - evaluate.py-1 B
    - perplexity.py-1 B
    - translation.py-1 B
    - stereoset.py-1 B
    - coreference.py-1 B
    - qa.py-1 B
  - ft
    - ft_hparams.py-1 B
    - ft_main.py-1 B
    - __init__.py-1 B
  - trace.py-1 B
  - generate_adapt_data.py-1 B
  - adapt_model.py-1 B
  - evaluate_model.py-1 B
  - utils
    - __init__.py-1 B
    - logit_lens.py-1 B
    - layer_stats.py-1 B
    - repr_tools.py-1 B
    - runningstats.py-1 B
    - globals.py-1 B
    - knowns.py-1 B
    - tok_dataset.py-1 B
    - notebooks_utils.py-1 B
    - constants.py-1 B
    - hparams.py-1 B
    - nethook.py-1 B
    - model_utils.py-1 B
    - generate.py-1 B
  - dama
    - compute_us_dama.py-1 B
    - __init__.py-1 B
    - dama_hparams.py-1 B
    - compute_v_dama.py-1 B
    - dama_main.py-1 B
  - notebooks
    - llama2_mt_adaptations_factual_stereotypical_traces.ipynb-1 B
    - globals.yml-1 B
    - averaged_casual_tracing_severing_mlps.ipynb-1 B
    - collect_results.ipynb-1 B
    - factual_stereotypical_traces_severing_mlps.ipynb-1 B