GECKOv2-ColBERT-NL-prod-v2

ColBERT retriever for GECKOv2 (nl), generated 2026-09-16 10:25.

Training configuration

Setting	Value
Base model	BAAI/bge-m3
Mode	all
Training strategy	single-round with cosine-sim negatives
Negative samples	10 per (query, positive) pair
nway	11 (1 positive + 10 negatives per step)
Learning rate	1e-05
Batch size	16 (accumsteps 1, 16 per forward pass)
Epochs	10
Search depth (k)	200
Mix ratio	0.0
Quantization (nbits)	4
Embedding model	sentence-transformers/all-MiniLM-L12-v2
Benchmark tables	included (production model)

Data

• Split: 1113 train / 148 validation questions, 214 train tables / 24 validation tables, 0 overlapping (should be 0)
• Collection: 169,224 documents

Rounds

Round	Negatives from	Triples	Steps	doc_maxlen	dim	Finished
1	cosine similarity of embedded descriptions	64,330	4,021	256	128	2026-09-11 15:57

Evaluation

Checkpoint	Dataset	EM	acc@k
colbertv2-dutch-all-nodes-production-dim128	complex_nl_test.jsonl	0.773	0.998 (@50)