Datasets:

---

joelniklaus

/

lextreme

like 22

P.S. We also have a pipeline for 17.5M Indian court decisions (Supreme Court + 25 High Courts, 1950–2026) from the AWS Open Data Registry (CC-BY-4.0). Once the HF dataset is published, Indian judgment prediction could be another LEXTREME task — a major underrepresented jurisdiction with both English and vernacular language coverage.

Hi Volodymyr,
Thanks for the message.

1. Yes, happy to look at it.
2. Retrieval seems out of scope for LEXTREME unfortunately. How do you represent the citation prediction task?
3. NLLP at EMNLP (https://2026.emnlp.org/calls/workshops/) could be a good venue. There might be a next edition of the DMAIL at ICDM workshop (https://dmail-workshop.github.io/DMAIL2025/) as well. And lastly there is ICAIL (https://site.smu.edu.sg/icail-2026), but the conference is soon, so you would need to wait for next year.

Hi Joel,

Sorry for the delayed follow-up — been heads-down running experiments. Two things:

1. Temporal drift paper (draft)

The paper is ready for your review. This is still a working draft, but the experiments and analysis are complete. Since my last message, we've expanded it significantly with two new experimental sections:

Section 5.5 — Continual Learning. Sequential fine-tuning across the three Ukrainian epochs. Key finding: chronological CL (pre-war → hybrid → full-scale) eliminates catastrophic forgetting for general XLM-R — pre-war F1 is fully retained (+1.8 to +6.2 pp) while full-scale gains +16.5 to +19.0 pp. Reverse-chronological CL causes severe forgetting (−12.2 to −14.3 pp). Legal-XLM-R forgets in both directions, suggesting legal-domain pretraining produces temporally fragile representations.

Section 5.6 — Cross-Jurisdictional Temporal Transfer. We use Swiss Judgment Prediction as a foreign-jurisdiction source, extending Cross-X to the temporal dimension. SJP pretraining lifts all Ukrainian epochs by +3 to +10 pp, but the forward degradation magnitude is unchanged (20.3 vs 21.3 pp). Temporal drift is orthogonal to cross-jurisdictional transfer.

The paper is 17 pages, 5 experiments, 6 tables, 5 figures. Draft PDF attached below. We'd particularly value your perspective on:

• The Legal-XLM-R finding (lower absolute but more temporally uniform -- a trade-off that could potentially be exploited)
• Whether temporal splits of SJP (2000--2020) could replicate the drift pattern in Swiss data -- this would make it a bilateral cross-jurisdictional temporal study
• Venue fit -- you mentioned NLLP at EMNLP, which seems ideal

2. Citation prediction task format

The most natural LEXTREME-compatible framing would be multi-label statute prediction (MLTC): given a court decision's facts section, predict which statutes (by article number) the court will cite in its reasoning. We'd limit the label set to the top-N most frequently cited statutes to keep it tractable.

The underlying data comes from our co-citation graph (overthelex/ua-court-citation-graph, 2.3M edges from 99.5M decisions). We can extract (facts_text → cited_statutes) pairs and format them as MLTC.

An additional dimension: the cited statute set evolves across temporal epochs (new Criminal Code articles post-2014 and post-2022), so this task would naturally combine citation prediction with temporal drift.

Would MLTC with top-N statutes work for LEXTREME, or do you have a different format in mind?

Best,
Volodymyr

Draft PDF: temporal-drift-legal-nlp-draft-2026-05-23.pdf

17 pages, compiles cleanly. Numbers are from actual training runs (MLflow), not placeholders. Still a draft — happy to iterate based on your feedback.

Hi Joel,

Reviving this: my last message (23 May) ended on an open question, so here is a concrete version of it.

Citation prediction as MLTC, drop-in for lextreme.py, same shape as the merged judgment subset:

_UKRAINIAN_STATUTE_CITATION = {
    "task_type": TaskType.MLTC,
    "hf_hub_name": "overthelex/ukrainian-statute-citation",
    "input_col": "text",        # facts section, untruncated
    "label_col": "labels",      # ["ЦК:625", "ЦПК:141", ...]
    "label_classes": [...],     # fixed top-N articles, N ~ 50
}

{"text": "<facts section>", "labels": ["ЦК:625", "ЦК:1166", "ЦПК:141"], "language": "uk"}

Input is the facts section, labels are {code}:{article} taken from the reasoning section, so no leakage from the same span. Fixed label set of the N most cited articles; out-of-vocabulary citations are dropped rather than collapsed into an "other" class.

Two options, your call:

(a) a single default config;
(b) three configs on the same epoch boundaries as the judgment subset (pre_war / hybrid_war / full_scale), which makes label drift observable, since new Criminal Code articles enter the label set post-2014 and post-2022. That is the version I find more interesting, at the cost of three configs instead of one.

Two things worth flagging before you decide. The per-decision pairs do not exist yet: ua-court-citation-graph only holds the aggregated co-citation graph (2.33M article-pair edges), so this needs a fresh export on our side. And the labels come from regex extraction, so extractor recall caps label quality; in follow-up work (arXiv:2606.00898) we found measured citation scores tracking oracle coverage rather than model behaviour. So I would ship extractor precision/recall, share of decisions with at least one in-vocabulary label, and mean labels per decision alongside the PR, and drop the task if those look bad.

Licence would be CC BY 4.0, matching the judgment subset.

Retrieval: understood, out of scope. ua-statute-retrieval stays standalone.

So: does MLTC with a fixed top-N label set work for you, and (a) or (b)? If yes, I will build the dataset, run the quality numbers, and open the lextreme.py PR.

Best,
Volodymyr