---
language:
  - uk
  - en
license: cc-by-4.0
pretty_name: Recency Probe 2026
size_categories:
  - n<1K
task_categories:
  - question-answering
  - multiple-choice
tags:
  - knowledge-cutoff
  - temporal
  - ukrainian
  - evaluation
configs:
  - config_name: default
    data_files:
      - split: test
        path: recency-probe-2026.parquet
---

# Recency Probe 2026

128 questions about facts that entered the record between **January and August
2026**, asked in Ukrainian and in English. A model trained before 2026 cannot
answer them from what it knows; a model that has read 2026 Ukrainian news can.

Built from
[`Goader/ukrainian-news-2026`](https://huggingface.co/datasets/Goader/ukrainian-news-2026)
— 429k articles from 23 Ukrainian national outlets. Every item is grounded in
quotes from that corpus, which ship with the item.

## What makes an item

Every candidate was put to a pre-2026 model
([`lapa-llm/lapa-12b-pt`](https://huggingface.co/lapa-llm/lapa-12b-pt), based on
`gemma-3-12b-pt`) in three framings, and kept **only if the model failed all
three**. So the questions are not merely recent — they are demonstrably outside a
2025-cutoff model.

Two kinds:

* **`stale`** (46 items) — someone or something *else* was the correct answer
  before 2025, and that predecessor is deliberately one of the four options.
  Answering means overwriting a belief, not filling a blank.
* **`gap`** (82 items) — no such predecessor. Four same-type alternatives with no
  privileged wrong answer.

Domains: international 48, war 24, government 17, law 11, politics 6, society 6,
culture 5, business 4, economy 4, sport 3. Answers are people (92), names (12),
organisations (10), places (8), numbers (4) and dates (2).

## Two formats

**Cloze, scored by likelihood** over four full-text continuations — the primary
read. No letters, no instruction-following, so a base model and an instruct model
can sit the same test and land in the same table.

```
Cloze:   З січня 2026 року посаду міністра оборони України обіймає ___
Options: Рустем Умєров · Степан Полторак · Олексій Резніков · Михайло Федоров
```

**Open short answer**, from `question_*` — meaningful after instruction tuning.

## Baseline

`lapa-llm/lapa-12b-pt`, cloze ranking over the four options. Chance is 0.250.

| slice | n | uk | en |
|---|---|---|---|
| all | 128 | 0.195 | 0.195 |
| `gap` | 82 | 0.305 | 0.305 |
| `stale` | 46 | **0.000** | **0.000** |

The zero on `stale` is not luck running out. Ranking the Ukrainian options, the
model picks the pre-2025 answer 67% of the time, another distractor 33%, and the
correct answer never — systematically wrong rather than randomly wrong.

## Fields

| column | description |
|---|---|
| `id` | item id, `pNNN` |
| `domain`, `answer_type` | subject area; person / org / place / name / number / date |
| `item_class` | `stale` or `gap`, as above |
| `question_uk`, `question_en` | the question, for open short answer |
| `cloze_uk`, `cloze_en` | sentence prefix; `cloze + " " + answer` is grammatical |
| `answer_uk`, `answer_en` | the correct answer |
| `aliases_uk`, `aliases_en` | other acceptable surface forms |
| `distractors_uk`, `distractors_en` | three wrong options |
| `choices_uk`, `choices_en` | answer + distractors, shuffled and fixed |
| `gold_uk`, `gold_en` | index of the answer in `choices_*` |
| `stale_uk` | the pre-2025 answer where one exists, else null; always among the distractors |
| `evidence` | the quotes the item was verified against — outlet, month, text |
| `n_articles`, `n_hosts` | how many articles and outlets state the fact |
| `in_train_mixin` | whether this fact's articles appear in the `train-mixin` config of the news dataset |

Every option set is checked mechanically: gold index correct, exactly three
distractors, no duplicate choices, the answer absent from its own cloze, evidence
spanning at least two outlets, no two items sharing an answer, and — because a
Ukrainian cloze governs case and gender — no option set where the correct answer
is the only grammatical fit.

## Use with lm-evaluation-harness

Four tasks ship in `lm-eval/`:

```bash
lm_eval --model hf --model_args pretrained=<checkpoint> \
        --include_path lm-eval --tasks recency_probe
```

`recency_probe_uk_cloze`, `recency_probe_en_cloze` (likelihood ranking, the
primary read), `recency_probe_uk_open`, `recency_probe_en_open` (generation), and
`recency_probe` to run all four.

```python
from datasets import load_dataset
ds = load_dataset("Goader/recency-probe-2026", split="test")
```

## Limits

* **128 items** — enough to compare checkpoints, small enough that a single
  fold's noise is visible. Report a confidence interval.
* **`gap` items sit at 0.305 against a 0.250 floor.** Their distractors are
  weaker than the `stale` half, where the pre-cutoff answer supplies one strong
  wrong option for free.
* **Some facts predate 2026** — a commander appointed in 2024, a spokesman in
  post since 2008. They are here because the control model failed them in five
  framings across two languages, which is the criterion; they measure "learnable
  from Ukrainian news" rather than strictly "2026-only".
* **The open tasks score by exact match** and ignore the alias lists; the harness
  needs a custom filter to use them. The cloze tasks are unaffected.
* Items were written and cross-checked by language models over human-written
  quotes; the quotes are in the dataset, so any item can be audited against its
  own evidence.