datause-classifier-multitask

Fine-tune of LiquidAI/LFM2.5-Encoder-230M with two heads: a page has_data gate and a document teratopic domain classifier.

Labels

30 teratopic sector themes (see encoder_labels.json).

Training

• base model: LiquidAI/LFM2.5-Encoder-230M
• epochs: 3
• learning rate: 2e-05
• max_length: 2048
• precision: bf16

Evaluation (holdout)

Gate (page has_data)

• threshold: 0.95
• precision: 0.9409
• recall: 0.8785
• F1: 0.9086
• % pages flagged negative: 0.4604

Domain (teratopic)

• micro F1: 0.5245
• macro F1: 0.3940
• mean average precision: 0.5031