author-set / README.md
saquiboye's picture
Upload 6 files
fb0998c verified
|
Raw History Blame Contribute Delete
1.24 kB

Training reads ntl_open_pool.csv (data.train_csv). Evaluation uses test.csv; data.test_fraction picks that share of complete prompt_ids (data.seed).

# ntl_open_pool.csv and test.csv in data/
uv run gemma-author-probe train

To regenerate the CSV from parquet: uv run gemma-author-probe export-data (defaults to data/ntl_open_pool.parquet).

Zust / CI pipelines: data/ntl_open_pool.csv and data/test.csv are in git. Parquet exports stay local (*.parquet gitignored). If only parquet is present, train.py can export the train CSV on startup. Optional: PROBE_PULSE_DATA_ROOT for an external data mount.

Parquet columns map to CSV as: id → id, prompt_group → prompt_id, text → text, model → model.

Training batches use ntl_open_pool.csv. With data.require_all_models: true (default), only prompts that contain all 18 models are batched. With false, any prompt with at least one model is eligible; if a sampled model id is missing on that prompt, a random model present on that prompt is used instead. A repeated model on the same prompt keeps the first row. Test eval still uses complete prompts (18 models each) from test.csv per data.test_fraction.