README.md

metadata

language:
  - nl
license: mit
library_name: transformers
tags:
  - bert
  - dutch
  - multi-head regression
  - text quality
  - sequence classification
model-index:
  - name: transformer_multi_head
    results:
      - task:
          type: text-regression
          name: Multi-Head Text Regression
        dataset:
          name: Proprietary Internal Text Dataset
          type: text
        metrics:
          - name: RMSE (delta_cola_to_final)
            type: rmse
            value: 0.14
          - name: R² (delta_cola_to_final)
            type: r2
            value: 0.4722
          - name: RMSE (delta_perplexity_to_final_large)
            type: rmse
            value: 0.0988
          - name: R² (delta_perplexity_to_final_large)
            type: r2
            value: 0.659
          - name: RMSE (iter_to_final_simplified)
            type: rmse
            value: 0.1231
          - name: R² (iter_to_final_simplified)
            type: r2
            value: 0.8587
          - name: RMSE (robbert_delta_blurb_to_final)
            type: rmse
            value: 0.1156
          - name: R² (robbert_delta_blurb_to_final)
            type: r2
            value: 0.7364
          - name: Mean RMSE (multi-head)
            type: rmse
            value: 0.1194
          - name: Aggregate RMSE (multi-head → final)
            type: rmse
            value: 0.0845
          - name: Aggregate R² (multi-head → final)
            type: r2
            value: 0.8146

transformer_multi_head

This is a multi-head transformer regression model based on GroNLP/bert-base-dutch-cased, fine-tuned to predict four separate text quality scores for Dutch texts.
The final aggregate metric re-computes a combined score from the four heads and compares it to the actual aggregate.

📈 Training & Evaluation

Epoch	Loss (Train)	Loss (Val)	RMSE (delta_cola)	R² (delta_cola)	RMSE (delta_perplexity)	R² (delta_perplexity)	RMSE (iter_to_final)	R² (iter_to_final)	RMSE (robbert_delta_blurb)	R² (robbert_delta_blurb)	Mean RMSE
1	0.0185	0.0152	0.1436	0.4447	0.1062	0.6066	0.1269	0.8500	0.1138	0.7446	0.1226
2	0.0141	0.0145	0.1400	0.4722	0.0988	0.6590	0.1231	0.8587	0.1156	0.7364	0.1194
3	0.0115	0.0146	0.1409	0.4656	0.0991	0.6571	0.1253	0.8537	0.1135	0.7458	0.1197
4	0.0094	0.0154	0.1468	0.4197	0.0985	0.6613	0.1297	0.8433	0.1164	0.7327	0.1228
5	0.0079	0.0154	0.1462	0.4246	0.1009	0.6444	0.1276	0.8482	0.1172	0.7291	0.1230

Final aggregate performance:

• Aggregate RMSE: 0.0845
• Aggregate R²: 0.8146

🧾 Notes

• This model uses four regression heads for: delta_cola_to_final, delta_perplexity_to_final_large, iter_to_final_simplified, and robbert_delta_blurb_to_final.
• The final performance aggregates the individual predictions back into a combined quality score for more robust quality measurement.
• Based on the Dutch BERT (GroNLP/bert-base-dutch-cased).

---