yhay81/reproducibility-audit-japanese-ai-agents

Japanese AI and Agent Evaluation Reproducibility Audit

Dataset description and intended users

This is a revision-pinned audit index for practitioners selecting Japanese AI resources and agent-evaluation benchmarks. Its 10 rows connect each Hub resource to a curation decision, license evidence, intended use, constraints, reproducibility status, and a known documentation or validation gap.

It is not a leaderboard, legal opinion, benchmark rerun, or endorsement. Scores from different cards are not made comparable by inclusion here.

Data files and schema

• data/resources.jsonl: one audit record per reviewed Model or Dataset.
• schema.json: JSON Schema for every row.
• audit-report.json: zero-drift Hub metadata check used for this release.
• release-manifest.json: source revision, row count, file sizes, and SHA-256 integrity hashes.
• CHANGELOG.md: version history and review state.
• notebooks/starter.ipynb: revision-pinned loading, decision summaries, attention filters, and evidence-integrity checks.

Every row carries the same 23 fields, all validated against schema.json:

Field	Type	Description
record_id	string	Stable unique identifier, <repo_type>:<repo_id>. Use it for joins; display titles are not identifiers.
schema_version	string	Version of the row schema (1.0.0 in this release).
repo_id	string	Hub repository ID of the audited resource, e.g. llm-jp/llm-jp-instructions.
repo_type	string	Audited repository kind: dataset or model.
collection	string	Curation track: japanese-ai, agent-evaluation, or contribution-hold.
decision	string	Curation decision, e.g. include, include-restricted, include-watch, include-documentation-gap, exclude-pending-clarification.
publisher	string	Organization or group that publishes the resource.
source_url	string	Canonical Hub URL of the audited resource.
reviewed_at	string	Timestamp when the human review baseline was recorded.
audited_at	string	Timestamp of the zero-drift Hub metadata audit for this release.
hf_cli_version	string	Official hf CLI version used for the read-only audit (1.26.0).
reviewed_revision	string	Exact Hub commit SHA of the resource at review time.
reviewed_last_modified	string	Hub lastModified timestamp observed at review time.
hub_license_tag	string	License tag shown in Hub metadata at the reviewed revision.
gating	string	Hub gating state at review time: false, auto, or manual.
license_evidence_url	string	Revision-pinned URL of the card or terms used as license evidence.
selection_rationale	string	Why this resource fills its evaluation role in the audit.
intended_use	string	Use that the audit record supports for this resource.
use_constraints	string	License- and terms-derived constraints to respect when using the resource.
reproducibility_status	string	Assessed reproducibility state, e.g. metadata-and-splits-documented, external-harness-required, restricted-card-audited.
known_gap	string	One known documentation or validation gap identified for the resource.
source_commit	string	Git commit of the private build repository that produced this release (build provenance, not a public reference).
ai_assisted	boolean	Always true: rows were drafted with AI assistance and human-reviewed.

Source data and licensing

Source records are public Hugging Face Hub metadata and maintainer-authored cards at exact revisions. Each row provides source_url and license_evidence_url. Upstream Models, Datasets, code, and referenced content retain their own licenses and terms; this Dataset does not relicense them.

The original audit structure and commentary in this repository are licensed under CC BY 4.0. Attribution should name Yusuke Hayashi and link this Dataset.

Collection and processing

The source manifest was reviewed at 2026-08-06T18:45:00+09:00. A read-only audit compares Hub SHA, last-modified time, license tags, and gating state with the baseline. The release builder refuses a non-zero-drift report, renders this card, emits JSONL and JSON Schema, and records hashes for the exact upload files.

The build repository is a private monorepo, so the commands below are maintainer-facing documentation: they run only inside that repository at commit 98b4aa58bb155cc76d6895066b34d451e449fd8c (recorded in every row as source_commit) and cannot be executed by third parties. Independent verification relies on the shipped machine-readable artifacts instead: validate data/resources.jsonl against schema.json, inspect the zero-drift audit-report.json, and check file integrity against the SHA-256 hashes in release-manifest.json. notebooks/starter.ipynb demonstrates these checks against a pinned Hub revision. Card-only documentation updates may postdate the release manifest, so the manifest README.md entry describes the card as of the data release; the data, schema, audit-report, and notebook hashes remain verifiable.

Maintainer build commands (private repository only):

uv run --with "huggingface-hub==1.26.0" python `
  projects/hugging-face-community/src/audit_hub_resources.py `
  --config projects/hugging-face-community/evidence/resources.toml `
  --output projects/hugging-face-community/outputs/hf-audit.json `
  --fail-on-drift

uv run python projects/hugging-face-community/src/build_audit_dataset.py `
  --audit-report projects/hugging-face-community/outputs/hf-audit.json `
  --output projects/hugging-face-community/outputs/huggingface-audit-dataset `
  --source-commit 98b4aa58bb155cc76d6895066b34d451e449fd8c

Quality, coverage, and missingness

The release validates an exact eight-file set, unique record identifiers, required schema fields, source-commit consistency, zero metadata drift, row count, and SHA-256 hashes. Narrative audit fields are human judgments grounded in the linked evidence; they are not automatically verified legal or scientific conclusions.

A clean ephemeral environment using datasets==5.0.1 loaded Hub revision a80df9e2c39356e8dd104eb73f4d875a28b08460 as 10 rows and 23 columns, with unique record_id values and the expected source commit. The starter notebook pins that immutable revision.

Coverage is deliberately narrow: Japanese text-generation resources and a progression from function calling through planning and open-ended tasks to executable software repair. Speech, vision, embeddings, domain-specific models, and general benchmark catalogs are outside scope.

Personal, sensitive, or restricted information

The Dataset contains no model weights, prompts, benchmark examples, answers, agent traces, personal data, or gated rows. Restricted resources such as AnswerCarefully and GAIA are represented only by public metadata, terms, and curation constraints.

Supported and prohibited uses

Supported uses include resource discovery, reproducibility reviews, license and gating triage, drift detection, and teaching revision-pinned ML practice. Do not use this Dataset to bypass upstream gates, infer rights absent from source terms, expose restricted examples, or claim that listed benchmark scores were independently reproduced.

Bias, risks, and limitations

Selection reflects a dated, intentionally small audit and may omit stronger or newer resources. Hub lastModified is an activity signal, not a quality score. License tags can be incomplete or inconsistent with card text. External harnesses, services, source repositories, and gated conditions can change after the review snapshot.

Update policy and version history

This is a snapshot audit, not a continuously monitored feed, and it carries no fixed update schedule. The most recent zero-drift metadata verification is the release audit recorded in audit-report.json (audited_at field; 2026-08-06 for this release). Re-checks and evidence reviews happen as the underlying resources change, with a review targeted within two weeks of a material license, gating, deprecation, or reproducibility change coming to the maintainer's attention. Rows are added only when they fill a missing evaluation role with inspectable evidence. Superseded, legally unclear, contaminated, or unreconstructable items are removed or explicitly quarantined rather than silently replaced. CHANGELOG.md records every released version.

Citation

@misc{hayashi2026reproducibilityauditjapaneseaiagents,
  author = {Hayashi, Yusuke},
  title  = {Japanese AI and Agent Evaluation Reproducibility Audit},
  year   = {2026},
  url    = {https://huggingface.co/datasets/yhay81/reproducibility-audit-japanese-ai-agents}
}

Plain-text form: Yusuke Hayashi (2026), Japanese AI and Agent Evaluation Reproducibility Audit, Hugging Face Dataset.

AI assistance and human verification

AI assistance was used for evidence organization, drafting, code generation, and consistency checks. A human authorized the work; machine-readable metadata and reproduction commands were verified against the recorded revisions. The audit does not imply independent verification of upstream benchmark claims.