Full fine-tuning of LiquidAI/LFM2.5-Encoder-230M for continuous Home Depot relevance. The model jointly encodes Query, Title, optional Brand, and Description. It pools the LFM BOS hidden state and predicts a score in [1, 3].

Validation RMSE: 0.426227
Validation MAE: 0.340394
Validation Spearman: 0.608588

This repository includes the fine-tuned backbone and regression head. Load a downloaded snapshot with LFM25EncoderRegressor.from_pretrained from LFM25/lfm25_crossencoder_model.py. The nested backbone requires trust_remote_code=True because it is an LFM2.5 Encoder model.