---
license: other
language: [hsb, dsb, de, en]
task_categories: [translation, text-generation]
pretty_name: LT3 WMT26 Sorbian reproduction datasets
tags: [wmt26, sorbian, upper-sorbian, lower-sorbian, low-resource, reproduction]
---

# LT3 at WMT26 (Sorbian): reproduction datasets

Companion **data** repo for the LT3 (Ghent University) WMT26 Sorbian system-description paper.
It carries the datasets needed to reproduce the submitted systems that **cannot** be fetched
automatically from an online source. Auto-fetchable inputs are deliberately excluded:
dictionaries, the organizer train/dev/test distribution, and raw public corpora already pinned
in the companion repo's `data/MANIFESTS/`.

- Model weights: https://huggingface.co/TomMoeras/wmt26-lt3-sorbian
- Code / pipeline: https://github.com/TomMoeras/wmt26-lt3-sorbian
- Fetchers: `data/fetch_pools.py` (pool) and `data/fetch_training_sets.py` (training) in the
  code repo pull this dataset and verify every file's sha256 against the frozen manifests
  (`data/pool_hashes.json`, `data/regeneration_recipe.json`).

## `pool/`: inference-time MT exemplar retrieval pool

The MT translation-memory pool used for fuzzy exemplar retrieval, in the per-direction
`<src>_<tgt>.src` / `.tgt` layout the inference stack expects. Consumed at **inference** time by
every FullStack system for fuzzy MT exemplar retrieval (k=3). Test-deduplicated: residual
test-source material is **0** in all six directions; the twelve files' sha256 match
`pool_hashes.json` exactly.

## `training/`: synthetic / derived training sets

One subdirectory per `regeneration_recipe.json` entry (sha256-matched to the recipe). Consumed
during model training.

| dir | what | seed / generator | consumed by |
|---|---|---|---|
| `mt_fuzzy_compfz` | **shipped** MT training rows (1,069,344, real-only, 0/1 compfz) | `src/data/{composition,build_training_set}.py` over the real parallel corpus, top-1 fuzzy@real | **submitted systems (base MT)** |
| `mr_messages` | MR training messages (28,147) | GSM8K (CC BY 4.0) + Hendrycks MATH (MIT), translated Sorbian question / English chain-of-thought | all systems (MR) |

## Provenance and licenses

Generators are ported and documented in the code repo under `training/generators/` and
`data/regeneration_recipe.json`; seed corpora and their licenses are enumerated in
`data/MANIFESTS/`. MR seeds are GSM8K (CC BY 4.0) and Hendrycks MATH (MIT). The MT training rows and the full (submission) pool derive from a mixture of
public Sorbian-German parallel corpora (OPUS, Leipzig CC BY-NC, WMT news-crawl research-use,
mtdata, and the WMT22 organizer MT parallel) plus model back-translation (the **published** pool
above carries the real parallel portion only); they are released here for
non-commercial research reproduction, inheriting the most restrictive terms of their sources
(non-commercial, attribution, research-use). Redistribution beyond research reproduction is not
granted.

## Deliberately excluded

- **Organizer competition data** (TUM-NLP distribution: train / dev / test, and any test-set
  derived rows). The `dev_fold` and `v10 training mix` recipe entries embed organizer dev rows
  in official eval formats and are therefore **regenerate-only** (rebuild with the code repo's
  `build_dev_fold.py` from the organizer distribution you fetch yourself).
- **GPL dictionary payloads** (soblex / dsb-spell content and the BK-tree pickles derived from
  them; the code repo's `setup.sh` builds those locally from the GPL sources).
- **SC/GC/QA synthetic** and the failed **QA official-format synth**: not archived as single
  files; regenerate via `training/generators/synth/` per the recipe.