Five character-level inflection transformer architectures (4 encoder + 4 decoder layers, embed dim 256, FFN 1024, 4 heads), trained on the 10L_90NL split; 12 runs per architecture (run id X_Y: X = data split / held-out test set, Y = training seed).

Layout

The repo mirrors the checkpoints/ subtree the probing pipeline expects:

checkpoints/
  vanilla/fixed_checkpoints/10L_90NL_<run>-models/checkpoint_best.pt   # fairseq 0.10.2 state_dict
  char_sep/seperate_char_checkpoints/10L_90NL_<run>-models/checkpoint_best.pt
  feature_onehot/independentfeature_fixed/10L_90NL_<run>.nll_0.0000.epoch_103
  feature_invariant/10L_90NL_<run>          # pickled transformer.Transformer
  feature_geometric/10L_90NL_<run>

Note: char_sep run 1_1 has checkpoint_last.pt instead of checkpoint_best.pt (no best checkpoint was saved for that run); the reproduction script handles this.

Loading

The vanilla and char_sep checkpoints are fairseq transformer checkpoints, but fairseq is not required: the probing repo loads them with a pure-PyTorch reimplementation of the fairseq forward pass (probing/extract_representations_vanilla.py, probing/extract_representations_char_sep.py). The other three architectures are pickled transformer.Transformer objects that need the model source code from the training repo (available from the authors).

To use with the probing pipeline:

hf download akki2825/probing-morphome-checkpoints --local-dir "$FEATURE_INFORMED_ROOT"

Citation

See the GitHub repository for citation information.