Ericu950/classical-swedish-citations

Classical → Swedish Citations

A dataset of labeled (Greek/Latin classical → Swedish) text pairs for studying classical reception, intertextual citation, and cross-lingual literary allusion. Each pair records a candidate citation surfaced by a retrieval + reranking pipeline over a Greek/Latin corpus and a Swedish-literature corpus, then judged by a large language model (Qwen3.6-27b) as either an unmistakable citation (TRUE) or a false positive (FALSE).

Quick start

from datasets import load_dataset

# Window-level pairs (5-sentence passages)
windows = load_dataset("Ericu950/classical-swedish-citations", "windows")
print(windows)
# DatasetDict with splits: greek_true, greek_false, latin_true, latin_false

# Sentence-level pairs (single-sentence matches)
sentences = load_dataset("Ericu950/classical-swedish-citations", "sentences")

A confirmed positive looks like this (Galatians 3:7 → Månsson's Swedish translation):

{
  "src_lang": "greek",
  "level": "sentence",
  "rerank": 0.9996,
  "sent_sim": 0.898,
  "src": {
    "author": "New Testament",
    "title": "Γαλάτας",
    "sentence_text": "Γινώσκετε ἄρα ὅτι οἱ ἐκ πίστεως, οὗτοι υἱοί εἰσιν Ἀβραάμ."
  },
  "tgt": {
    "book": "MånssonF_Rättfärdiggörelsen1",
    "sentence_text": "Veten därför, att de som äro av tron, de äro Abrahams barn."
  },
  "llm_answer": "TRUE",
  "llm_p_true": 0.998
}

Configurations

The dataset has two configurations, corresponding to two granularities at which matches were judged:

Config	Granularity	Use case
windows	5-sentence passages	Thematic citations, broader allusions
sentences	Single sentences	Tight translations, direct quotations

Each configuration has four splits (greek_true, greek_false, latin_true, latin_false), giving classical-language × LLM-verdict.

Statistics

Approximate counts (run len(load_dataset(...)["split"]) for exact numbers):

	Greek TRUE	Greek FALSE	Latin TRUE	Latin FALSE
Windows	~1,500	~40,000	~320	~6,900
Sentences	(varies)	(varies)	(varies)	(varies)

Schema

Windows config

rank          int       Pipeline rank of this candidate (1 = highest composite score)
level         str       "window"
src_lang      str       "greek" or "latin"
score         float     Composite z-scored ranking signal
rerank        float     Cross-encoder probability of citation [0,1]
win_sim       float     Window-window cosine similarity
sent_sim      float     Sentence-sentence cosine similarity (secondary)
src_density   float     Mean cosine to background source pool (lower = rarer)
tgt_density   float     Mean cosine to background target pool
src_rarity    float     1 - percentile_rank(src_density), clipped
tgt_rarity    float     1 - percentile_rank(tgt_density), clipped
src_idx       int       Position in the source embedding cache
tgt_idx       int       Position in the target embedding cache
src_bin       int       sentence_index // 5 (used for dedup)
tgt_bin       int       sentence_index // 5
src           dict      {author, title, book, file, lang, lang_conf, sentence_index, window_text}
tgt           dict      same fields, target side
llm_answer    str       "TRUE" or "FALSE"
llm_p_true    float     P(TRUE) from first-token logprobs

Sentences config

Same as windows, except:

• level is "sentence"
• src_bin / tgt_bin are replaced by src_sent_i / tgt_sent_i (exact sentence indices, since bin size is 1)
• src and tgt blocks contain an additional sentence_text field (the focal sentence used for matching; window_text is the surrounding context)

Construction methodology

This dataset was produced by an LLM-judge bootstrap pipeline rather than human annotation. The process:

1. Synthetic seed: a multilingual bi-encoder (BAAI/bge-m3) was fine-tuned on ~119k Greek/Latin → Swedish synthetic translation triples produced by Qwen 3.6-27b.
2. Retrieval: the fine-tuned bi-encoder retrieved top-K Swedish target windows / sentences for each Greek and Latin source window / sentence in the corpus.
3. Reranking: a fine-tuned cross-encoder (BAAI/bge-reranker-v2-m3), finedtuned on the synthetic paralell data, classified each candidate pair.
4. Composite scoring + filtering: pipeline outputs were filtered by a composite score combining cosine similarity, reranker probability, and distinctiveness (rarity vs. corpus background).
5. LLM judgment: each surviving candidate was judged by Qwen3.6-27B-Instruct with a strict philological prompt. Pairs judged as unmistakable translations, citations, paraphrases, or direct allusions were marked TRUE; the rest, FALSE.
6. Iterative improvement: cross-encoder was further finetuned on retrived allusions, before this set of allusions were produced and judged.

For sentence-level judgments, the LLM was given both the focal sentence and the surrounding 5-sentence window to enable narrative-context veto of otherwise plausible-looking sentence matches.

Intended use

• Training and evaluating cross-lingual citation detection models
• Studying classical reception in 19th-20th century Swedish literature
• Benchmarking semantic similarity / reranker models on classical languages
• Hard-negative mining for retrieval models

The FALSE examples are particularly valuable as hard negatives — these are pairs that a strong retrieval + reranking pipeline mistook for citations, then a strong LLM rejected. They reflect realistic failure modes of citation detection systems.

Limitations

• Not human-annotated. Labels reflect LLM judgment under a strict prompt. Spot-checking suggests precision on TRUE is high, but biases inherent to the LLM judge may be present.
• Coverage is bounded by retrieval recall. Citations that the bi-encoder failed to retrieve in the top-K candidates are not in the dataset, either as positives or as negatives.
• Domain skew. The Swedish corpus over-represents 19th- and early 20th-century literature (Strindberg, Söderberg, Rydberg, etc.).
• Length sensitivity. Sentence-level positives can be sensitive to short, generic phrasing despite the strict prompt.

License

CC BY 4.0. Underlying source texts are in the public domain (classical texts from Perseus; Swedish texts from Litteraturbanken). Cite both this dataset and the original sources when using.

Citation

@dataset{classical_swedish_citations_2026,
  author = {Eric Cullhed},
  title  = {Classical → Swedish Citations: A bootstrapped dataset of
            Greek/Latin classical citations in Swedish literature},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Ericu950/classical-swedish-citations}
}