mirror of
https://github.com/guillaumemeyer/watermarks-remover.git
synced 2026-08-22 13:11:57 +02:00
guillaumemeyer
dfad1d55f1
Tracks the research/ arXiv-v1 paper kit (per PR #174's gitignore plan) and closes the environment gaps needed to actually run the 3,500-cell study: - detect_text_watermark.py / multilingual_gen.py: new --torch-dtype (auto|fp32|bf16). bf16 is ~8x faster for opt-1.3b CPU cached decode (0.5 vs 4 s/token on aarch64) and keeps same-config gen/detect consistent. - run_experiments.py: --quality-python (defaults to repo .venv-quality, else MarkLLM venv with a warning); --seeds/--prompts now limit the generate loop instead of only the plan; workers force --torch-dtype bf16; the 'none' control attack is a passthrough (was 'unknown attack'). - evaluate_quality.py: BERTScore on roberta-large (deberta-xlarge-mnli crashes under transformers>=5 in set_truncation_and_padding; WATERMARKS_BERTSCORE_MODEL override). - Makefile: research-check target (documented but missing). - .env.example: MARKLLM_DIR naming (code reads MARKLLM_DIR, not WATERMARKS_MARKLLM_DIR). - research/pins-quality.txt: pinned quality-venv freeze (protocol §7). - tests: quality-python plumbing, CLI subset, bf16 worker flag, none attack, multilingual fake_load signature. Validation: make research-check (78 tests) and make lint green; a restricted end-to-end run (generate -> attack -> detect -> evaluate -> report) passes with real metrics.