Nini0la/edgeimci-beta0-1k-multitask-enriched-2258-v1

EdgeIMCI Beta0-1K Multitask Enriched 2258

Dataset summary

This is the public, message-normalized publication of the dataset used to fine-tune the selected EdgeIMCI 4B-alpha-second research checkpoint from the pinned Qwen/Qwen3-4B base model. It combines structured extraction, routing, bounded clinical-response, project self-knowledge, scope-safety, and proposition/negation examples for the EdgeIMCI sick-child assessment workflow.

The dataset is research evidence for the ADTC 2026 submission. It is not clinical guidance, a diagnostic dataset, or authorization to deploy the model in patient care.

Exact identity

manifest.json and checksums.sha256 are the machine-readable publication records.

Splits

The publication preserves the pre-existing training and validation assignments. No examples were shuffled, resampled, or repartitioned for publication.

There is no public or private TEST split in this release.

Load the Parquet publication with:

from datasets import load_dataset

records = load_dataset(
    "Nini0la/edgeimci-beta0-1k-multitask-enriched-2258-v1",
    revision="fac66e746466130604ea3eaa37c6e38fc1dd0b13",
)

This SHA is the immutable first publication commit containing the dataset files. Submission metadata must use an exact commit SHA, never main.

Record schema

Each row contains:

• —component_id: source component within the multitask release.
• —example_id: stable example identifier.
• —leakage_group_id: grouping key used by the split-overlap controls.
• —lineage_parent_ids: source-parent identifiers.
• —messages: ordered system, user, and assistant chat messages used for SFT.
• —mode: EXTRACTION or FREE_FORM.
• —partition: frozen TRAIN or VALIDATION assignment.
• —review_tag: review state where present.
• —semantic_parent_id and source_case_id: lineage identifiers.
• —evaluation_suite, expected_route, and validation_status: validation metadata; null for rows where not applicable.

Training composition

The release consists of the frozen 1,831-row multitask release plus 427 enrichment records:

The 1,831-row base release itself contains retained Beta0-1K extraction records plus free-form reasoning, project self-knowledge, scope/safety, proposition/negation, extraction controls, and routing companions. Full component counts are in manifest.json and provenance/baseline_manifest.json.

Construction and provenance

The clinical target for each Alpha 3 enrichment record was computed before language generation by the local deterministic EdgeIMCI oracle. GPT-5.4 was given facts and wording instructions; it was not authorized to choose classifications or management targets. Delta 1 targets inherit pinned project facts. Delta 2 records implement the bounded scope/refusal and urgent-referral policy.

The source expansion audit checked JSON schema validity, unique IDs and prompts, baseline-prompt conflicts, validation leakage groups, inherited Delta 1 targets, Alpha 3 oracle binding, output shape, and numerical-fact preservation. These automated checks do not constitute independent clinical review.

The original 427 candidate-schema records are retained in provenance/addition_candidates.jsonl. The publication builder converts them to the same chat-message schema as the base release without rewriting user or assistant message text. The builder and source manifests are included under provenance/.

Publication normalization

The frozen training receipt identifies the exact internal TRAIN file by SHA-256. That internal normalized file was not retained in the reachable repository state. This publication therefore reconstructs the training view from the hash-verified 1,831-row base release and the hash-verified 427-record enrichment source.

• —User and assistant message text is taken directly from those source records.
• —The publication adds nullable columns so both splits expose one Hugging Face/Arrow schema.
• —The published TRAIN JSONL/Parquet hashes are publication-artifact hashes and are not represented as the frozen internal TRAIN-file hash.
• —The published VALIDATION JSONL remains byte-identical to the frozen validation file.

This distinction is recorded explicitly in manifest.json rather than treating differently serialized files as byte-identical.

Leakage and split controls

The source release reported no detected overlap between TRAIN and VALIDATION under its bounded checks for transitive lineage, leakage group, normalized prompt, normalized target, and normalized evidence. Packaging also rejects direct leakage-group overlap and duplicate TRAIN example_id values.

This is not proof of arbitrary semantic independence. Shared labels and clinical concepts can occur across splits.

Intended uses

Appropriate uses include:

• —reproducing or auditing the selected EdgeIMCI research fine-tuning run;
• —studying bounded structured extraction and routing;
• —inspecting the ADTC submission's training-data provenance; and
• —non-clinical research on small-model behavior in a constrained workflow.

Out-of-scope uses

Do not use this dataset or a model trained on it as:

• —an autonomous diagnostic or treatment system;
• —a replacement for qualified clinical judgment;
• —evidence of clinical validation or regulatory approval;
• —a source of patient-specific medical advice; or
• —a general-purpose pediatric or emergency-care dataset.

Data characteristics and limitations

• —Records are curated, generated, or transformed research examples rather than representative field data.
• —Language coverage is primarily English and does not establish performance across Nigerian languages or real PHC documentation styles.
• —Clinical and source-governance review remains pending.
• —Validation contains structured extraction/routing examples and no free-form validation examples.
• —The published set does not establish safety, calibration, fairness, or clinical effectiveness.

Personal and sensitive information

The release is constructed from synthetic or curated scenarios and is not intended to contain real patient records or direct personal identifiers. That design claim is not a substitute for an independent privacy audit.

License and access

license: other means no general reuse license is asserted for this release. Public access is provided for ADTC judging and research inspection. Users are responsible for verifying any rights required for reuse of upstream clinical guidance and generated derivatives.

Public availability does not change the dataset's research-only status or authorize clinical deployment.

Citation

Until a formal citation is issued, cite the immutable Hugging Face repository commit together with:

EdgeIMCI Beta0-1K Multitask Enriched 2258, release beta0_1k_multitask_enriched_2258_v1, 2026.