ABench75

A compact bilingual (Russian + English) benchmark for instruction-following and short-answer reasoning across five domains: math, logic, code, cybersecurity, general.

• Total tasks: 375 (75 per domain)
• Public split: 300 tasks (data/tasks_public.jsonl)
• Private holdout: 75 tasks (15 per domain, not published) — for honest tracking
• Languages: ~59% English / ~41% Russian
• Scoring: regex match on the final answer (re field, case-insensitive)

Canary: ABENCH75_CANARY_8f3c1d9a. If you find this string in a training corpus, the benchmark has leaked — treat contaminated scores accordingly.

Format

Each line of data/tasks_public.jsonl is one JSON object:

{"id": "math_ru_0", "lang": "ru", "domain": "math",
 "q": "Сколько будет 74 умножить на 56? Ответь только числом.",
 "re": "\\b4144\\b"}

• id — stable identifier
• lang — ru | en
• domain — math | logic | code | cyber | general
• q — the prompt
• re — Python regex the generated answer must match (search, ignore case)

Usage

pip install -r requirements.txt   # torch, transformers, peft (optional)
python run_eval.py --model <hf-model-or-path> --tasks data/tasks_public.jsonl
# optional: LoRA adapter + reasoning-level tag
python run_eval.py --model ./merged --adapter ./lora --effort high --batch 8

The runner prints overall accuracy and breakdowns by domain and language.

Scoring is intentionally simple: the answer must match re. Models that emit reasoning should put the final answer in plain text (the runner strips common <think>…</think> wrappers first).

Results (public split, 300 tasks)

model	params	overall	math	logic	code	cyber	general
Velary 1.2 Flash (medium)	1.2B	82.7%	97%	49%	85%	96%	85%
Velary 1.2 Flash (high)	1.2B	81.6%	96%	43%	85%	96%	88%
Velary Mini (no-think)	0.35B	50.7%	32%	39%	67%	55%	61%

Numbers are from the public 300-task split; the private 75-task holdout is reported separately by the maintainers.

Design notes & limitations

• Short-answer regex scoring favours direct final answers; it is not a judge for long-form quality.
• 75 tasks per domain is enough for a quick signal, not for fine-grained ranking; treat differences of a few points as noise.
• Categories are deliberately broad; see data/tasks_public.jsonl for coverage.

Split policy

The public/private split is stratified by domain (60 public / 15 private per domain). The private holdout is kept out of this repository to reduce contamination and to provide a cleaner progress signal.

License

• Data (data/tasks_public.jsonl): CC BY 4.0.
• Code (run_eval.py, tooling): MIT — see LICENSE.