metadata
license: mit
pretty_name: ALIEN data and results
ALIEN — data and results
Data, model adapters and experiment outputs for ALIEN: Aligned Entropy Head for Improving Uncertainty
Estimation of LLMs (submitted results + revision experiments). Code: the ALIEN repository,
revision/ folder.
Layout
| folder | content |
|---|---|
results/ |
Published runs used for the paper tables: results_on_embeddings/grid_* (ALIEN, its variants and ablations; best_run.pickle = test scores + metric_df with the selected hyper-parameters), baseline_results_on_embeddings_rde/ (MD, RMD, MDM, RDE), probes_ds_with_lr_hyper/ (linear / attention-pooling probes), full-fine-tuning and prompted-LLM blocks |
data/embeddings/ |
Pooled train/val/test features, logits, targets and the classification head per task model: <model>_results_ds_<seed>_train_gird_full/LoraConfig_<dataset>_state.pickle |
data/models/ |
LoRA adapters of the task models: <model>_runs_seeds/<Dataset>/<seed>/ |
revision_runs/heldout/head_seeds/ |
Base model seed 0: ALIEN, ConfidNet, last-layer linear probe re-trained with up to 20 head seeds (hyper-parameters fixed per model × dataset) |
revision_runs/heldout/head_seeds_base{1..4}/ |
Same for base model seeds 1–4 (5 head seeds); *_own_hp/ = runs with each base seed's own hyper-parameters |
revision_runs/heldout/alien, confidnet, embeddings_heldout |
Held-out-error protocol pilot (heads trained on 85 % of the validation split) |
revision_runs/probes_seeds/ |
Linear + attention-pooling probes (last / mid / begin layer) with head seeds: model_<m>/dataset_<D>/base_<k>/<probe>_layer_<L>/head_seed_<s>.pt, lr.json = learning rate per probe |
revision_runs/outputs/ |
Aggregated tables (per-seed metrics, paired tests, calibration, selective risk) |
Reproduce the paper metrics
With the ALIEN repository checked out and this dataset downloaded to DATA:
python revision/scripts/reproduce_paper_tables.py --results-dir DATA/results --out-dir out --seed-dirs seed_0
python revision/scripts/compare_with_paper.py --results-dir DATA/results --out-dir out
python revision/scripts/make_selection_tables.py --results-dir DATA/results --out-dir out --tables-dir out/tables
Seed-based statistics: revision/scripts/summarize_nested.py --work DATA/revision_runs/heldout --results-dir DATA/results.