PaulBurca2005/ro-retrieval-triplets-late-interaction

Romanian retrieval triplets (late-interaction hard negatives)

(anchor, positive, negative) triplets for training a Romanian retriever, generated with Gemma 3 27B over a merged Romanian corpus and mined for hard negatives with late interaction (MaxSim).

A BM25 variant of the same triplets is at PaulBurca2005/ro-retrieval-triplets. The two are close to disjoint - over the queries both contain, mean Jaccard overlap between their negative sets is 0.053 and 73% share no negative at all - so training on the union exposes a model to both failure modes.

The column layout follows alina0195/ro-msmarco-divided, with one column added.

column
anchor	the query
positive	the document the query was generated from
negative	one hard negative, mined by late interaction (MaxSim)
query_source	the upstream dataset or outlet the query comes from

One row per (query, positive, negative) pair, so a query with four negatives is four rows repeating the anchor and the positive.

How it was built

The positive is not retrieved, it is known: every query was generated from a document, so that document is the positive by construction. Retrieval is used only to mine negatives.

Retrieval is late interaction: one embedding per token rather than one per document, scored with MaxSim - for every query token, the best matching document token, summed. Two stages, because MaxSim over a whole corpus per query is not affordable: mean-pooled vectors give a shortlist, MaxSim reranks it. A candidate becomes a negative if it scores inside a band relative to the positive's own MaxSim score and is not the positive's own story.

This is not a trained ColBERT. There is no ColBERT checkpoint for Romanian, so the token embeddings come from a multilingual sentence-transformer trained for mean-pooled similarity, not for late interaction. It earns its place by ranking differently from BM25, not by being a faithful ColBERT.

The corpus is deduplicated by a near-duplicate key first, boilerplate is dropped, and self-referential queries ("what does the article say…") are removed - identically in both variants, so the two are comparable.

Sources

source	rows	share
realitatea	8 399	10.4%
BlackKakapo/recipes-ro	6 757	8.3%
protv	6 732	8.3%
libertatea	6 568	8.1%
readerbench/ro-text-summarization (alephnews)	6 417	7.9%
zf	6 011	7.4%
evz	5 844	7.2%
readerbench/ro-stories	5 712	7.1%
mediafax	5 625	7.0%
digi24	5 367	6.6%
readerbench/ro-text-summarization (digi24)	5 259	6.5%
aleph	4 875	6.0%
cotidianul	3 781	4.7%
adevarul	3 321	4.1%
readerbench/ro-text-summarization (mediafax)	260	0.3%

Known limitations

Read these before training on it.

1. Residual false negatives. The guards remove candidates that answer the query better than the positive does, not every candidate that happens to answer it. A second article about the same event, worded differently, can still appear as a negative. The builder also records candidates it judged to be the positive's own story — that judgement was measured at roughly 70% precision, because lexical overlap sees shared topic, not shared event: it cannot separate two different earthquakes, or two different articles about the same person. Those are recorded for a reranker to re-judge, not treated as verified positives.
2. Source imbalance. The corpus is a merge of news outlets, folk tales, recipes and summarisation corpora with very different sizes and registers. query_source exists so this can be weighted or filtered; the distribution above is not uniform.
3. Machine-generated queries. The queries are Gemma 3 27B output over the source documents. They were filtered but not human-verified.
4. No stemming. The BM25 tokenizer folds diacritics but does not stem, and Romanian is heavily inflected, so lexical matching is weaker than it could be.

Splits

eval, test, train.

Where splits exist they are grouped by the positive's duplicate group, so a query and a re-published copy of its positive never land on opposite sides.