DRU-RE EntityPair TwoHead
Arabic relation extraction model based on NAMAA-Space/AraModernBert-Base-V1.0.
Architecture
The marked sentence is encoded by AraModernBERT. The unchanged classifier explicitly extracts:
- the contextual CLS vector,
- the contextual opening subject-marker vector,
- the contextual opening object-marker vector,
- the absolute subject/object vector difference,
- the elementwise subject/object vector product,
- a 42-dimensional directional OneHot entity-type vector.
These are concatenated into 5H + 42 dimensions and compressed through a
1024 -> 512 pair MLP.
The shared pair representation feeds two heads:
- binary relation-existence head: no relation vs relation exists;
- 40-way positive-relation head: relation type, with no
no_relation class.
For no_relation rows, only the existence loss is used. The positive-relation
head is trained only on positive rows.
Threshold
The existence threshold is tuned on labeled val.jsonl to maximize positive
micro F1. The threshold used for the official prediction was
0.0100 from auto_tuned_on_labeled_val.
Main files
model/pytorch_model.binmodel/modeling_entity_pair_two_head.pyconfigs/run_config.json (includes the exact base encoder and resized tokenizer size)configs/architecture_config.jsonartifacts/existence_threshold_sweep.csvartifacts/submission.zipartifacts/predictions.txt
