jensjepsen/esperanto-orca-math
Esperanto Orca-Math Chain-of-thought math reasoning examples in Esperanto, translated from microsoft/orca-math-word-problems-200k via jensjepsen/eo-mt-v5b (MarianMT en→eo). v5b's occasional operator-drop artifacts (e.g. 6 * 9 → 6 9) were repaired post-hoc by detecting A B = C patterns where A * B == C and re-inserting the *. Format: {"messages": [{"role": "user", "content": Q_eo}, {"role": "assistant", "content": A_eo}]} Files train.jsonl (128,689 rows) —… See the full description on the dataset page: https://huggingface.co/datasets/jensjepsen/esperanto-orca-math.
Esperanto Orca-Math
Chain-of-thought math reasoning examples in Esperanto, translated from microsoft/orca-math-word-problems-200k via jensjepsen/eo-mt-v5b (MarianMT en→eo). v5b's occasional operator-drop artifacts (e.g. 6 * 9 → 6 9) were repaired post-hoc by detecting A B = C patterns where A * B == C and re-inserting the *.
Format: {"messages": [{"role": "user", "content": Q_eo}, {"role": "assistant", "content": A_eo}]}
Files
- `train.jsonl` (128,689 rows) — filtered for Q+A ≤ 512 tokens in the morpheme tokenizer of
jensjepsen/eo-base-44k. Use this if you're training a small model with a tight context window. - `train_full.jsonl` (197,849 rows) — all rows with valid Q and A, including the ones too long for 512 tokens. Use this if your model has a wider context.
Both files also include gold (the GSM-style final numeric answer) and, for the long-row variant, filter_reason (which budget gate they tripped).
Provenance
Translation pipeline: mt/scripts/translate_traces.py → mt/scripts/fix_op_drops.py in jensjepsen/small-esperanto-llm.