README.md

metadata

license: cc-by-4.0
language:
  - hy
  - en
task_categories:
  - text-generation
  - question-answering
tags:
  - arxiv:2609.03350
  - armenian
  - math
  - science
  - reasoning
  - machine-translated
  - verified
  - parallel-corpus
size_categories:
  - 100K<n<1M
pretty_name: ArmSTEM
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train/*

ArmSTEM

ArmSTEM is a corpus of 373K verified machine-translated mathematics and science problems, 324K of them with step-by-step solutions, released as parallel English–Armenian (~311M Armenian + ~124M English tokens, Gemma tokenizer). To our knowledge it is the first Armenian STEM corpus with worked solutions at training scale.

Every item passed a verification pipeline whose central gate is blind re-solving: an independent model solves the Armenian problem and must reproduce the original gold answer exactly. Translation preserved not just fluency but correctness.

In our experiments (see paper), mixing just 6% ArmSTEM into Armenian continued pretraining reversed catastrophic forgetting — the adapted model ended above its base on a six-task Armenian suite (+9.7pp on Belebele) instead of 21 points below it.

Sister releases: COPA-AI/armweb and COPA-AI/arm-gemma-e4b.

Composition (v1)

Source (English)	License	Items accepted	Accept rate
GSM8K	MIT	7,404	99.1%
AceReason-Math	CC-BY-4.0	48,584	98.0%
OpenScience	CC-BY-4.0	264,266	97.4%
OpenScienceReasoning-2	CC-BY-4.0	52,653	91.5%

Translation of the full source pools is complete (372,907 unique verified pairs). AceReason-Math provides final answers without worked solutions, so its 48,584 items carry cot_en = null and cot_hy = the answer; the other 324,323 items include step-by-step solutions on both sides. An automated adequacy audit (GPT-5.5, 300 stratified items, 1–5 scale) rates 100% of re-solve-verified and 92.7% of solver-limited items as meaning-preserving (means 4.65 / 4.39); the audit file ships in stats/. The audit judge (GPT-5.5) also serves as the escalation translator for a small fraction of items. For a complete verification of the translated samples, two native Armenian speakers independently assessed a 300-item sample for logical coherence and solution correctness, rating 299/300 valid with identical verdicts on every item (Cohen's κ = 1.0). Blind re-solve marks 10.9% of accepted math and 27.6% of science items solver_limited (the English original also resists the solver, so the gate certifies problem-statement integrity only). A harder competition-math tranche will ship as v1.1.

Fields: id, src (source dataset), question_en, cot_en (solution; null for AceReason-Math), question_hy, cot_hy, gold, answer_type (numeric/mc/freeform), solver_limited (bool, see gate G2 below), translator, attempts. Rejected items and per-stage rejection statistics ship in stats/, together with stats/cpt_training_subset_ids.txt, the 104,630 ids of the CPT training subset that appear verbatim in this release (the remaining 5,255 OpenScience items of that subset were superseded by revised translations before release).

Pipeline

1. English-side decontamination: 13-gram scan against English benchmark origins (above all MMLU-Pro test, the source of ArmBench's largest column). Contaminated items are dropped before translation.
2. Placeholder masking: numbers, LaTeX spans, and the question/solution separator are replaced with indexed placeholder tokens (⟦N1⟧, ⟦EQ2⟧, …) before translation and restored afterward — number/notation corruption, the dominant MT failure mode for math, is eliminated structurally.
3. Translation: Gemini-3.1-flash-lite (selected over GPT-5.5 by a fixed-protocol 200-item A/B), with two feedback-guided repair rounds and escalation to a stronger translator on repeated failure.
4. Gates: G0 placeholder integrity (every token exactly once, no stray digits) → G1 language ID (GlotLID) → G2 blind re-solve: o4-mini solves the Armenian problem; the final answer must match the gold exactly (numeric/MC). On mismatch, a control re-solves the English original — if that also fails, the item is solver-limited, not mistranslated, and is kept with a solver_limited tag. Freeform answers (<5%) use a three-model judge panel, 2/3 majority.
5. Armenian-side decontamination: accepted translations are scanned against the full ArmBench-LLM item set (both gates rejected real collisions in production).

Per-stage rejection statistics ship with the corpus (stats/).

License, attribution, statement of changes

Each subset inherits its source license (MIT for GSM8K; CC-BY-4.0 for the NVIDIA-released sets). Per the CC-BY requirements: the material was machine-translated to Eastern Armenian with automated verification; original sources are credited above and per-item in the src field. OpenScience/OpenScienceReasoning-2 are Qwen-generated synthetic data: the data license is CC-BY-4.0, and per the upstream note, models trained on them may inherit Qwen-license considerations.

Decontamination notice: ArmSTEM is decontaminated against both its English benchmark origins and Armenian targets (ArmBench-LLM, including MMLU-Pro-Hy). It is safe to train on and evaluate on those benchmarks. m-MMLU-hy and ARC-hy were not decontamination targets; a post-hoc 13-gram scan found 4 of the 372,907 pairs sharing a 13-gram with their items.

Citation

@article{arakelyan2026armweb,
  title  = {From Zero to Hero: An Open LLM Ecosystem for Armenian},
  author = {Arakelyan, Erik and Avetisyan, Khatun and Davtyan, Meri and Grigoryan, Heghine and Khachatryan, Nane and Shahsuvaryan, Hayk and Sergoyan, Henrik and Martirosyan, Vahan},
  year   = {2026},
  journal = {arXiv preprint arXiv:2609.03350}
}

Please also cite the source datasets (GSM8K, AceReason-Math, OpenScience, OpenScienceReasoning-2) when using the corresponding subsets.