Model Card for coref-llm-CRAC26-multilingual

This model is a fine-tuned version of google/gemma-3-27b-it. It has been trained using TRL.

CRAC 2026 Shared Task — Test Set Results

Dataset	F1 Score
cs_pcedt	72.79
hu_korkor	60.54
en_gum	73.88
la_coreflat	33.56
cs_pdt	76.31
no_bokmaalnarc	81.19
cs_pdtsc	70.89
fr_democrat	56.76
ca_ancora	78.00
pl_pcc	78.35
cu_proiel	61.32
nl_openboek	77.80
hi_hdtb	73.74
hu_szegedkoref	62.81
ko_ecmt	68.10
hbo_ptnk	75.65
lt_lcc	60.45
tr_itcc	68.39
grc_proiel	74.11
fr_ancor	77.33
en_litbank	81.56
no_nynorsknarc	79.35
es_ancora	78.30
fr_litbankfr	73.34
de_potsdamcc	70.87
en_fantasycoref	78.47
ru_rucor	81.99
Average	71.33

Evaluation on the official CRAC 2026 shared task test set.

Training procedure

This model was trained with SFT.

Framework versions

• PEFT 0.18.1
• TRL: 0.29.0
• Transformers: 5.2.0
• Pytorch: 2.10.0+cu126
• Datasets: 4.6.1
• Tokenizers: 0.22.2