22ling1/meddies-embedding-testbatch-v0

Meddies multilingual retrieval queries

This package combines the reviewed bilingual query work from Round 3 with the final 1,500-query cross-lingual output from Round 6. It contains 1,700 query–passage records derived from 310 passages.

Dataset composition

Source	Records	Languages	Review state
Round 3	200	English and Vietnamese	196 approved core positives; 4 held for partial support
Round 6	1,500	17 query languages	Pending fluent-language and entity review

Round 3 contains 100 information needs expressed once in English and once in Vietnamese. Round 6 contains five distinct information needs for each of 300 passages, with four secondary-focus queries and one main-focus query per passage.

Intended use

The records support multilingual and cross-lingual retrieval experiments, embedding evaluation, query–passage ranking, and quality-review workflows. Each record includes the query, its positive passage, an answer sketch, and exact evidence spans.

Use training_eligible == true to select the 196 reviewed Round 3 core-positive rows. Round 6 rows remain available for evaluation and review, but are deliberately marked ineligible until fluent-language and entity-form review is completed.

Loading

from datasets import load_dataset

dataset = load_dataset("YOUR_ACCOUNT/YOUR_DATASET_NAME", split="combined")
reviewed = dataset.filter(lambda row: row["training_eligible"])

To load the local Parquet file:

from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files="data/combined.parquet",
    split="train",
)

Core fields

Field	Description
record_id	Unique stable record identifier prefixed by source round
intent_id	Identifier shared by related query formulations
query	Natural-language retrieval query
query_language	Query language code
passage	Unchanged positive passage
passage_language	Passage language code
pair_type	cross_lingual or same_language
focus	main or secondary information need
expected_answer	Short answer sketch supported by the passage
evidence_spans	Exact source substrings supporting the answer
passage_sha256	SHA-256 of the UTF-8 passage text
training_eligible	Whether the row is approved for the core training pool
review_status	Current review state

Round-specific annotations are preserved in nullable fields. Round 3 supplies lexical relationship, starting-knowledge, naturalness, overlap, and answerability annotations. Round 6 supplies canonical-entity and generation-review metadata.

Validation

The package was checked for a uniform schema, unique record IDs, valid passage hashes, exact evidence spans, correct same-language/cross-lingual labels, and duplicate query–passage pairs. The Parquet file was also read back and compared field-for-field with all 1,700 source records. See validation_report.json, parquet_validation.json, and stats.json.

Limitations

• Round 6 translations and localized entity forms have not received independent native-speaker review.
• Round 3 review was unblinded and performed in shared generation context.
• Two Round 3 intents are retained for traceability despite partial support; their four language rows have review_status: held_partial_support and training_eligible: false.
• The data contains no hard negatives and does not itself demonstrate downstream embedding gains.
• Medical, nutritional, administrative, and time-sensitive statements are grounded only against the supplied passages and were not independently fact-checked.

Before publishing the dataset publicly, confirm that the source passages may be redistributed and add the appropriate license metadata. No license has been inferred in this package.

Provenance

The dataset follows the project's query-generation principles: standalone information needs, exact source evidence, natural localized entity forms, explicit uncertainty preservation, and separate metadata for identity tracking. Original Round 3 and Round 6 files remain unchanged.