amis-unified-multilabel-xlmr

AMIS Unified Multilabel XLM-R Classifier

This is the best checkpoint from the AMIS unified multilabel classifier training run. It predicts the thematic or commodity relevance of a news chunk across 10 labels in one forward pass.

Labels

Training Dataset

The model was trained on the unified dataset faodl/amis-agri-unified. The dataset is created by loading the source binary datasets above, normalizing their text and binary labels, grouping duplicate texts within each topic, outer-joining all topics by exact text, and using -1 for labels that were not annotated for a given topic. The split used for this run is the exact split produced by seed=42.

Split Label Coverage

Split Label Totals

Training Parameters

Validation Thresholds

These thresholds were selected independently per label on the validation split by maximizing positive-label F1.

Validation Metrics

Macro F1 at threshold 0.5: 0.8129
Macro F1 with per-label best-F1 thresholds: 0.8280

Test Metrics

The tuned test metrics use thresholds selected on the validation split.

Macro F1 at threshold 0.5: 0.8160
Macro F1 with validation-selected per-label thresholds: 0.8181

Inference

Use sigmoid over the model logits. For default predictions, use threshold 0.5 for every label. For tuned predictions, use the per-label thresholds shown above in label order.