AstheralDataset

15,032 examples: 5,000 sourced translation pairs and 10,032 original AI-authored instruction-response examples.

Contents

Target	Train	Validation	Test	Total
German	900	50	50	1,000
French	900	50	50	1,000
Spanish	900	50	50	1,000
Italian	900	50	50	1,000
Portuguese	900	50	50	1,000
Total	4,500	250	250	5,000

Record fields

Each row contains a unique SHA-256 ID, English source text, translated target text, source and target language codes, a translation instruction, and the original dataset/config/split/revision/row index.

Curation

Text is drawn from real parallel sentences in OPUS-100. Selection preserves the original train, validation and test split membership. All 10,000 text values are globally unique after Unicode normalization, case folding, punctuation removal and whitespace normalization. Source-target identity, short fragments, markup and placeholder patterns are filtered. No model-generated translations are included.

The complete build script, exact source revision, split counts, rejection counts and Parquet checksums are published alongside the data.

Load

from datasets import load_dataset

dataset = load_dataset("SEE7/AstheralDataset")
example = dataset["train"][0]
print(example["source_text"])
print(example["target_text"])

Provenance and attribution

Source: Helsinki-NLP/opus-100, revision 805090dc28bf78897da9641cdf08b61287580df9. OPUS-100: Zhang et al. (2020), “Improving Massively Multilingual Neural Machine Translation and Zero-Shot Translation.” Original parallel text comes from the OPUS collection. The original source corpus terms remain applicable.

Curated dataset repository: SEE7/AstheralDataset.

Original examples — configuration: original

10,000 newly generated examples have been added to the original 32: 2,508 coding, 2,508 short stories, 2,508 sentence exercises and 2,508 safety-alignment examples.

Category	Train	Validation	Test	Total
Coding	2,256	126	126	2,508
Stories	2,256	126	126	2,508
Sentences	2,256	126	126	2,508
Safety alignment	2,256	126	126	2,508
Original total	9,024	504	504	10,032

Safety responses are generated from explicit authored safety guidance and checked for required safety positions. The guidance is published in safety_guidance.json.

The new coding, stories and sentence content was generated specifically for SEE7 with Qwen2.5-1.5B-Instruct; safety responses use Qwen2.5-7B-Instruct. Generation uses authored task, subject, setting and style combinations. No additional external example dataset was imported. Provenance is recorded in each row. The original translation pairs remain attributed to OPUS-100.

Every new instruction and response passed normalized exact-duplicate checks against the existing dataset and all accepted new content. Truncated responses, placeholder patterns, incomplete code and malformed Python were rejected. The 2,500 new coding answers passed AST parsing and compilation checks. The original eight coding examples retain their recorded execution tests.

Use the original configuration to browse the additions. original_examples.jsonl contains all 10,032 instruction-response records. original_validation.json records generation settings, counts, rejection reasons, validation methods and file checksums.

from datasets import load_dataset
original = load_dataset("SEE7/AstheralDataset", "original")
translations = load_dataset("SEE7/AstheralDataset", "default")

Read a sample immediately · Browse all original examples

Final release checks: 10,000 new examples, 2,500 per category, 30,064 globally unique normalized instruction/response and translation text values. Full dataset: 15,032 rows. The eight original executed coding tests are preserved in legacy_original_32_tests.json.

Safety-alignment access

Read safety examples immediately · Download all 2,508 safety rows as JSONL · Download CSV

These are category exports of the existing original configuration, not additional dataset rows.