ACROS: Sense Representations are Inducible Interfaces

Hugging Face collection (models/checkpoints): https://huggingface.co/collections/jcblaise/acros

Paper: https://arxiv.org/abs/2605.28669

Install

python -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

Data setup

bash scripts/download_data.sh

Most scripts also download required datasets lazily if not already cached.

Repository layout

• run_sense_init.py, run_sense_induction.py: ACROS sense initialization and induction training
• run_adaptation.py: SENSiA adaptation
• run_post_adaptation_eval.py: intrinsic/downstream post-adaptation evaluation
• run_flores_ppl.py: FLORES sentence-level PPL evaluation
• evals/run_wsd.py: unified WSD entrypoint
• evals/run_steering.py: unified CoInCo steering entrypoint
• evals/run_generation.py: unified XL-Sum generation/report entrypoint
• evals/run_significance.py: significance/CI entrypoint

Main experiment commands

1) WSD

python evals/run_wsd.py acros -- \
  --model_name_or_path <HF_MODEL_OR_LOCAL_CHECKPOINT> \
  --output_json eval_logs/wsd/gloss_activation_target_lemma_colon_raganato_all.json

Baselines:

python evals/run_wsd.py gloss_lm -- \
  --model_name_or_path HuggingFaceTB/SmolLM2-360M \
  --output_json eval_logs/wsd/gloss_lm_base_smollm2_360m_raganato_all.json

python evals/run_wsd.py mfs -- \
  --strategy wordnet_first \
  --output_json eval_logs/wsd/mfs_wordnet_raganato_all.json

2) CoInCo lexical steering

Build cases:

python evals/run_steering.py build_coinco -- \
  --output_json eval_logs/coinco_lexsub/coinco_test_cases.json

Run target-best ACROS steering:

python evals/run_steering.py targetbest -- \
  --model_name_or_path <HF_MODEL_OR_LOCAL_CHECKPOINT> \
  --cases_json eval_logs/coinco_lexsub/coinco_test_cases.json \
  --output_json eval_logs/coinco_lexsub/coinco_targetbest.json

Run non-oracle self top-k selector:

python evals/run_steering.py self_topk -- \
  --model_name_or_path <HF_MODEL_OR_LOCAL_CHECKPOINT> \
  --cases_json eval_logs/coinco_lexsub/coinco_test_cases.json \
  --output_json eval_logs/coinco_lexsub/coinco_self_topk.json

3) SENSiA adaptation + intrinsic evaluation

python run_adaptation.py \
  --model_name_or_path <HF_MODEL_OR_LOCAL_CHECKPOINT> \
  --output_dir outputs/adapt_eng_ind

python run_post_adaptation_eval.py \
  --model_name_or_path outputs/adapt_eng_ind \
  --output_dir eval_logs/post_adapt

python run_flores_ppl.py \
  --model_name_or_path outputs/adapt_eng_ind \
  --output_json eval_logs/flores_ppl/adapt_eng_ind_sentence.json

4) XL-Sum generation (paper-style generation metrics)

python evals/run_generation.py xlsum -- \
  --model_name_or_path <HF_MODEL_OR_LOCAL_CHECKPOINT> \
  --lang ind \
  --split test \
  --output_jsonl eval_logs/xlsum_full/jsonl/acros_ind_fulltest.jsonl

python evals/run_generation.py xlsum_report -- \
  --input_glob 'eval_logs/xlsum_full/jsonl/*.jsonl' \
  --output_json eval_logs/xlsum_full/report.json