Felixbrk/robbert-2023-dutch-large-multi-score-text-only

transformer_multi_head_robbert2023

This is a multi-head transformer regression model based on DTAI-KULeuven/robbert-2023-dutch-large, fine-tuned to predict four separate text quality scores for Dutch texts.

The final aggregate metric recomputes a combined score from the four heads and compares it to the actual aggregate, providing robust quality tracking.

---

📈 Training & Evaluation

Epoch	Train Loss	Val Loss	RMSE (delta_cola_to_final)	R² (delta_cola_to_final)	RMSE (delta_perplexity_to_final_large)	R² (delta_perplexity_to_final_large)	RMSE (iter_to_final_simplified)	R² (iter_to_final_simplified)	RMSE (robbert_delta_blurb_to_final)	R² (robbert_delta_blurb_to_final)	Mean RMSE
1	0.0271	0.0179	0.1478	0.4119	0.1247	0.4569	0.1408	0.8151	0.1192	0.7196	0.1331
2	0.0191	0.0168	0.1495	0.3980	0.1038	0.6239	0.1368	0.8255	0.1233	0.7004	0.1283
3	0.0166	0.0151	0.1471	0.4172	0.0994	0.6548	0.1278	0.8477	0.1125	0.7503	0.1217
4	0.0148	0.0145	0.1421	0.4557	0.0969	0.6720	0.1254	0.8535	0.1131	0.7478	0.1194

Final aggregate performance
✅ Aggregate RMSE: 0.0860
✅ Aggregate R²: 0.8078

---

🧾 Notes

• This model uses four parallel regression heads for:
  • delta_cola_to_final
  • delta_perplexity_to_final_large
  • iter_to_final_simplified
  • robbert_delta_blurb_to_final
• The final evaluation recomputes a single aggregate score for robust performance tracking.
• Fine-tuned on a proprietary dataset of Dutch text variations.
• Base: RobBERT 2023 Large (DTAI-KULeuven/robbert-2023-dutch-large).

---

✅ Usage Example

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("DTAI-KULeuven/robbert-2023-dutch-large")
model = AutoModelForSequenceClassification.from_pretrained(
    "your-org/transformer_multi_head_robbert2023",
    num_labels=4,
    problem_type="regression"
)

inputs = tokenizer("Dit is een voorbeeldzin in het Nederlands.", return_tensors="pt")
outputs = model(**inputs)
print(outputs.logits)  # shape: [batch_size, 4] → 4 predicted scores