creator-reranker-v2

Fine-tuned reranker trained with preference learning (RewardTrainer / Bradley-Terry loss) on top of creator-reranker-v1.

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model     = AutoModelForSequenceClassification.from_pretrained("YourHFUsername/creator-reranker-v2")
tokenizer = AutoTokenizer.from_pretrained("YourHFUsername/creator-reranker-v2")

def score(query, fact):
    inputs = tokenizer(query, fact, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        return model(**inputs).logits.squeeze().item()

print(score("works in fashion and beauty", "beauty influencer, collaborates with luxury brands"))

Training

• Base model: BAAI/bge-reranker-base → fine-tuned → creator-reranker-v1 → preference training → creator-reranker-v2
• Training method: RewardTrainer (Bradley-Terry pairwise loss)
• Pairs: ~3,755 preference pairs (gap ≥ 0.3) derived from scored (query, fact, label) data
• Best checkpoint: epoch 1 (early stopping)
• Label range: 0.0 (undesirable) → 1.0 (desirable)