2610gnam05/polypulse-deberta-v3-base-demand

PolyPulse — DeBERTa-v3-base

Fine-tuned from microsoft/deberta-v3-base to classify proposed Polymarket questions as Low, Medium or High trading-demand categories. The target is historical USD trading volume, not the probability of the underlying event occurring.

• Dataset: SII-WANGZJ/Polymarket_data (same prepared 50K/3K/3K chronological splits used for the model comparison).
• Training: 1 epoch, mild class weighting, learning rate 1e-5.
• Held-out tuned test accuracy: 0.7146666666666667.
• Held-out tuned test macro-F1: 0.6115108829621564.
• Important: load polypulse_config.json and add logit_biases to model logits before softmax/argmax; the default pipeline() prediction alone is not the tuned result.
• Input text: Question: {question} [SEP] Event: {event_title} (use the same format as training).
• max_length and historical USD thresholds are also in polypulse_config.json.

Limitations: Accuracy varies across demand classes; model scores are not guaranteed calibrated probabilities. The model is not investment advice.