esherialabs/saferide-gemma-4-e2b-v058-original-419806-training-data

SafeRide Synthetic Bilingual Safety Guidance Dataset v0.5.8

This research and development dataset contains synthetic English and Kiswahili chat conversations. It was designed to help a language model practice cautious, agency-preserving safety guidance, useful refusal behavior, and responses that avoid inventing facts. It contains no real survivor reports or production records. The frozen dataset is publicly available under Creative Commons Attribution 4.0 International (CC BY 4.0). Public availability does not establish real-world safety, effectiveness, or production readiness.

Key facts

Field	Value
Unique rows	1,904
Weighted training examples	1,992
Languages	English and Kiswahili, 952 unique rows each
Content source	Synthetic only
Available split	Train only
Access	Public research dataset
Primary artifact SHA-256	669835b5f680b2198cf06d8c23a02d2e9aba2acba816c6283a0cdbbf8586a15e

Intended uses

Researchers and developers may use this dataset to:

• reproduce the training input for the linked SafeRide v0.5.8 PEFT adapter;
• study bilingual, synthetic conversational safety data under controlled research conditions;
• audit the documented data composition, weighting, and training lineage; and
• develop and evaluate safer response patterns before any real-world testing.

Out-of-scope and responsible-use boundaries

CC BY 4.0 permits reuse, adaptation, and redistribution with attribution. The following are product and safety boundaries, not additional copyright-license restrictions. Public availability must not be treated as authorization for:

• use as evidence that a model is safe, effective, culturally appropriate, or useful to survivors in the real world;
• direct legal, medical, counselling, emergency, investigative, or eligibility decisions;
• training systems for surveillance, coercion, profiling, or decisions about an identifiable person; or
• adding real survivor narratives, evidence, audio, transcripts, locations, credentials, production logs, or other personal or confidential data.

Reusers must provide attribution, link to CC BY 4.0, indicate changes, and must not imply endorsement by Esheria Ventures Limited, SafeRide, or UNICEF.

Loading the dataset

The repository is public. No Hugging Face credential is required to load the immutable training revision.

from datasets import load_dataset

repo_id = "esherialabs/saferide-gemma-4-e2b-v058-original-419806-training-data"
artifact_revision = "ab43518babcf6255fddf7ae0087f7ce78a84a707"
training_data = load_dataset(
    repo_id,
    "original-419806-weighted",
    split="train",
    revision=artifact_revision,
)

This example is syntax-checked. The immutable JSONL files and their public metadata are separately checksum-verified at the revision shown above.

Configurations and weighting

The repository exposes two configurations:

• original-419806-unique contains 1,904 deduplicated conversations and is the clearest configuration for inspection and analysis.
• original-419806-weighted contains the exact 1,992-example sequence used for training. It repeats a small set of targeted synthetic mitigation examples so those patterns have greater influence during the single training epoch.

Use the weighted configuration to reproduce training. Its additional 88 entries are repeated examples, not 88 new conversations. Use the unique configuration when calculating corpus composition or reviewing distinct rows.

Dataset structure

Each JSON Lines record is one synthetic conversation with these top-level fields:

Field	Description
id	Stable row identifier.
datasetId	Source dataset version.
schema, schemaVersion	Record-format identifiers.
candidateId	Synthetic candidate identifier used during curation.
stage	Curation stage recorded by the source pipeline.
split	Always train in this release.
messages	Ordered chat messages with role and content fields.
metadata	Language, scenario, risk, category, strategy, and review metadata.
authoring	How and when the synthetic row was produced and attested.

The observed role sequences are:

• system → user → assistant for 1,376 single-turn rows; and
• system → user → assistant → user → assistant for 528 multi-turn rows.

The metadata object contains appState, conversationForm, generatorVersion, language, longResponseReason, primaryCategory, prohibitedDataScreen, responseSkeletonId, responseStrategy, reviewLedgerRefs, reviewStatus, reviewableContentSha256, riskLevel, scenarioFamilyId, secondaryTags, semanticClusterId, sourceKind, sourcePolicyRefs, systemPromptSha256, and userGoalCode.

The authoring object contains authorIdentityRef, authoringPromptSha256, configurationSha256, createdAt, method, scenarioFamilyId, status, syntheticOnlyAttested, termsAssessmentRef, toolId, and toolRevision. Hashes and identifiers support audit and deduplication. They do not replace a review of the released row content or its associated policies.

System-message disclosure warning

Every row contains a system-role message. The contents are intentionally not reproduced in this card, but they are part of the public JSONL files. SafeRide completed its prompt, policy, privacy, security, and intellectual-property disclosure review for this release on August 13, 2026. Reusers remain responsible for assessing the messages and resulting behavior for their own context.

Composition statistics

All counts below describe the 1,904 unique rows, not the weighted training sequence.

Language and conversation form

Dimension	Count
English	952
Kiswahili	952
Single-turn conversations	1,376
Multi-turn conversations	528

Synthetic risk labels

These are authoring labels used to shape dataset coverage. They are not assessments of a real person or incident.

Label	Count
Critical	302
High	450
Medium	424
Low	424
Not assigned	304

High-level categories

Category	Count
Coercion	160
Emergency guidance	160
Fabricated-information avoidance	160
Jailbreak and instruction extraction	160
Legal boundaries	160
Medical boundaries	160
No-new-facts behavior	160
Privacy	160
Product-truth boundaries	160
Tone and agency	160
Not assigned	304

Creation and curation

The corpus combines several synthetic-only source lineages. Of the unique rows, 1,600 are repository-authored synthetic conversations, 120 are repository-pipeline-authored synthetic conversations, 112 are deterministic synthetic mitigation rows, and 72 are deterministic human-authored synthetic mitigation rows.

Source files were accepted only after deterministic schema, role, split, identity, length, lineage, and checksum checks. Duplicate identifiers were rejected, and the final unique file was byte-bound before the weighted training sequence was produced. Review controls differed by source lineage; this card does not claim that every row received an independent human language or domain review.

Development prompts, evaluation holdouts, and post-training continuation data are excluded. No held-out material was used for optimizer updates.

Personal and sensitive information

The dataset was designed and attested as synthetic-only. It intentionally excludes real survivor records, evidence, exact locations, credentials, private communications, and production telemetry. Synthetic-only status reduces but does not eliminate disclosure and misuse risk. The project completed its privacy and content disclosure review for this public release on August 13, 2026; future versions require their own review.

Biases and limitations

• Synthetic conversations do not demonstrate real-world usefulness, safety, or outcomes for survivors.
• Kiswahili coverage has not completed independent language review.
• The corpus cannot represent the full cultural, regional, disability, socioeconomic, legal, medical, or safeguarding diversity of real situations.
• Sheng is not included or supported by this release.
• The dataset has only a training split and is not itself an evaluation benchmark.
• Repeated mitigation examples intentionally alter training frequency and can increase both desired behavior and unwanted over-refusal.
• Results depend on the base model, chat template, system policy, decoding settings, and downstream safeguards; the dataset alone cannot establish a safe application.

Versioning and maintenance

The training files are frozen at immutable Hugging Face revision ab43518babcf6255fddf7ae0087f7ce78a84a707. Documentation-only commits are recorded separately and do not change that artifact revision. Any row-level change requires a new dataset version, new hashes, and new model evaluation.

The repository is public. Public visibility does not change the immutable data revision, license, limitations, or need for language, privacy, security, and provenance review.

Historical freeze metadata

The immutable training revision was originally frozen while the repository was private. Its hash-bound MANIFEST.json and docs/GOVERNANCE.md therefore retain historical private-only wording. Those files remain unchanged for auditability. PUBLIC_RELEASE.md records the later approval that supersedes only that access status; it does not rewrite the frozen data, hashes, lineage, or limitations.

License, citation, maintainers, and contact

The frozen dataset is licensed by Esheria Ventures Limited under the Creative Commons Attribution 4.0 International license. The license permits sharing and adaptation, including commercial reuse, when users provide appropriate credit, link to the license, and indicate changes. It does not grant rights to third-party material, trademarks, confidential evaluation content, real survivor data, or any material explicitly excluded from this release.

Suggested attribution:

SafeRide Synthetic Bilingual Safety Guidance Dataset v0.5.8, Esheria Ventures Limited, CC BY 4.0, immutable revision ab43518babcf6255fddf7ae0087f7ce78a84a707.

The maintainer organization is Esheria Ventures Limited under the esherialabs namespace. The public maintainer and security contact is Franklin Sagini at sagini@esheria.ai.

Esheria Ventures Limited gratefully acknowledges financial support provided for this Project by the UNICEF Innovation Fund. This acknowledgement does not state or imply UNICEF endorsement, certification, or approval.

Related artifacts

• PEFT adapter
• LiteRT-LM mobile package
• SafeRide website
• Android v0.5.8 testing preview and checksum
• Canonical open-source repository

Technical provenance and integrity

The main facts above are intended for general readers. The following bindings support exact reproduction and audit.

Evidence	Value
Dataset repository	esherialabs/saferide-gemma-4-e2b-v058-original-419806-training-data
Immutable artifact revision	ab43518babcf6255fddf7ae0087f7ce78a84a707
Unique JSONL SHA-256	669835b5f680b2198cf06d8c23a02d2e9aba2acba816c6283a0cdbbf8586a15e
Weighted JSONL SHA-256	b6fd044f9e7854d358200288b195787fc9ed6e8eea52925eea9aa0d48783689e
Manifest SHA-256	1cbb5931f418f3c63ad5c671ab7617f54ed679fc745ea96f7543511bdc638076
Dataset license	CC-BY-4.0
Source evidence commit	229ac0ab1eebefda6dc623b948503a087206dd35

The immutable manifest records source lineage, exclusions, model links, and privacy assertions. The checksum ledger binds every file in the frozen release. README/model-card revisions and hashes are tracked separately from these immutable data bytes.