license: agpl-3.0
language:

• en
• ms
  tags:
• constitutional-ai
• governed-execution
• bahasa-melayu
• penang-loghat
• arifos
• i-arif
• sft
• dpo
• qlora
• sovereign-ai
• f1-f13
• agentic-ai
• mcp
  pretty_name: I-ARIF CANON
  size_categories:
• 1K<n<10K
  task_categories:
• text-generation
• question-answering
  format:
• jsonl
  modality:
• text
  library:
• datasets
• pandas
• mlcroissant

configs:

• config_name: agentic_rl
  data_files:
  • split: train
    path: agentic-trajectories-rl.jsonl
    splits:
  • name: train
    num_examples: 50
• config_name: a2b_heldout
  data_files:
  • split: train
    path: i-arif-a2b-heldout.jsonl
    splits:
  • name: train
    num_examples: 100
• config_name: agentic
  data_files:
  • split: train
    path: i-arif-agentic-trajectories.jsonl
    splits:
  • name: train
    num_examples: 14
• config_name: dpo
  data_files:
  • split: train
    path: i-arif-dpo.jsonl
    splits:
  • name: train
    num_examples: 96
• config_name: redteam
  data_files:
  • split: train
    path: i-arif-redteam-eval.jsonl
    splits:
  • name: train
    num_examples: 100
• config_name: sft
  default: true
  data_files:
  • split: train
    path: i-arif-sft.jsonl
    splits:
  • name: train
    num_examples: 1322
• config_name: val
  data_files:
  • split: validation
    path: i-arif-val.jsonl
    splits:
  • name: validation
    num_examples: 17

🗺️ Position in I-ARIF Governance Stack

This dataset is part of the arifOS constitutional governance training-and-evaluation pipeline — a closed-loop alignment substrate.

Total: 1,122 downloads across 8 datasets · verified 2026-08-11
🔗 Full stack: https://huggingface.co/ariffazil · 🌐 Federation: https://arif-fazil.com

I-ARIF CANON — Governed Execution Substrate Training Corpus

NOT an LLM dataset. This is the constitutional training corpus for Project I-ARIF:
a Governed Execution Substrate that wraps AI model outputs with runtime floor enforcement
(F1-F13), action classification (ART), policy gates (OPA/Cedar), witness receipts, and
immutable audit (VAULT999).

What This Corpus Trains

I-ARIF shifts governance from textual system prompts (soft, bypassable) to runtime
policy gates and action classifiers (hard, code-enforced). This corpus teaches a base
model to operate WITHIN constitutional constraints as native behavior.

Corpus Composition (1,435 pairs)

Source-of-Truth Graph

This corpus is a composite of the six upstream constitutional datasets:

SFT (1,322 pairs):
  ariffazil/AAA  ── constitutional genome (1,024 canon + 258 nusantara + 26 floor) ──┐
  trajectories   ── agentic execution traces (14 pairs)                             ──┤
DPO (96 pairs):                                                                     │
  ariffazil/BBB  ── guardrail failure baseline (55 pairs)                           ──┼──►  I-ARIF-CANON
  ariffazil/CCC  ── kernel mediation proof (7 pairs)                                ──┤    (1,435 active pairs)
  ariffazil/DDD  ── register sensitivity (34 pairs)                                 ──┤    (+ 100 held-out)
VAL (17 pairs):                                                                     │
  ariffazil/EEE  ── kernel spine audit (5 VAL probes)                               ──┤
  ariffazil/FFF  ── federation fitness gate (12 VAL probes)                          ──┘
                                                                        ▼
                                                               arifOS QLoRA Phase 2

Language Distribution

• English: 88% (canonical governance definitions)
• Bahasa Melayu / Penang loghat: 12% (Nusantara semantic substrate)
• Target: 30% BM for Phase 2 QLoRA training

Audit Status — Updated 2026-08-06

PARTIAL SEAL → EXECUTION BOUNDARY. All static code claims verifiable on disk.
Live telemetry claims cite probe commands. Benchmark claims pending Phase 2 QLoRA run.

The breakthrough is not that the model becomes incapable of error.
The breakthrough is that error is no longer allowed to travel naked.

Source SHA-256 (manifest.json, 2026-07-29 forge)

• i-arif-sft.jsonl: 9ace2816bd2258cbb791bf99259f4a58f85b1fcc7627de27ced6269b241d93c1
• i-arif-dpo.jsonl: fc9904086cd172f22935515b4d697ad6795611425f29320f7c9fc1b01abbe22e
• i-arif-val.jsonl: 80ecc082180eb6b30d16708c8ae6093e34a1696a2c9f5f28d2e4d5784892270e

Baseline (DeepSeek V4 Pro)

• eta (Governance Compliance): 0.375
• Target after QLoRA: >= 0.888
• Gap to close: 0.513
• Probe breakdown:
  • EEE kernel spine: 2/3 correct
  • FFF model fitness: 1/3 correct
  • a2b benchmark scenarios: 0/2 (held-out AssetOpsBench — not a primary gate)

Readiness Gap (2026-08-06)

The corpus is NOT yet production-ready for training I-ARIF. Two gaps remain:

1. Scale floor. 1,308 SFT pairs is below the threshold where constitutional
   behavior is baked in rather than style-transferred. Target: ≥50k pairs.
2. No agentic trajectories. I-ARIF claims agentic behavior but the SFT set is
   single-turn Q&A. Multi-turn rollouts with tool calls, ART classification, and
   witness-receipt generation are absent. VAL set now has 17 clean F1-F13 probes but
   needs expansion with multi-turn gating scenarios.

Resolved (2026-08-06, Gap 2): Domain contamination removed. 100 AssetOpsBench
scenarios moved to i-arif-a2b-heldout.jsonl. VAL gate now tests F1-F13 governance
only (5 EEE kernel spine + 12 FFF model fitness).

In progress (2026-08-06, Gap 3): Agentic trajectory pipeline deployed — 3 real multi-turn
RL trajectories (HEALTH_AUDIT, CONSTITUTIONAL_REFUSAL, FQ_RECOVERY) with per-floor reward
scores across all 9 constitutional floors. Each trajectory encodes full 6-plane loop with
real tool calls, AUTH failure recovery, and deterministic F1-F13 scoring. Target: 50+ trajectories.

Done (2026-08-06, Gap 4): Constitutional reward spec written — F1-F13 as deterministic
scoring functions (binary + continuous), not LLM-as-judge. Aggregation: weighted sum with
thresholds SEAL_READY (0.85) / NEEDS_IMPROVEMENT (0.60) / UNSAFE (0.40).

Phase 2 QLoRA run depends on closing these gaps. Until then, this is v0.1 seed.

Files

• i-arif-sft.jsonl — Supervised fine-tuning pairs
• i-arif-dpo.jsonl — Direct preference optimization pairs
• i-arif-val.jsonl — Validation gating probes
• manifest.json — Provenance and composition manifest
• w3_baseline_deepseek_v4_pro.json — Pre-training baseline scores
• i-arif-agentic-trajectories.jsonl — SFT agentic descriptions (14 seed pairs)
• agentic-trajectories-rl.jsonl — RL training trajectories (50 forged multi-turn tool-call scenario templates with per-floor reward scores)
• constitutional-reward-spec.json — RLAIF reward function spec (F1-F13 as deterministic scorers, not LLM judges)
• i-arif-redteam-eval.jsonl — Red-team adversarial eval: 100 probes across 8 attack categories (command injection, sovereignty bypass, register manipulation, authority confusion, jailbreak, tool misuse, seal tampering, gradual escalation). All 13 floors covered.
• build_iarif_corpus.py — Corpus synthesis pipeline (v2: --force, --no-overwrite, --dry-run safety gates)
• SFT_SCALE_PIPELINE.md — Gap 1 design spec: 1,322→50k pair scale architecture, 5 categories, template-driven synthetic generation
• AUDIT-READY-DOSSIER.md — Full architectural dossier (audit-ready v2)

Cross-Dataset Pointer

Federated Integration — Updated 2026-10-04

This dated documentation update separates the governance specification, candidate neural assessors, local operational evidence, and the immutable dataset payload. It does not announce a trained model or production readiness. Earlier dated corpus counts and experiments remain historical snapshots.

Governance scope and provenance

The i-ARIF Blueprint defines a constitutional contract for a Governed State Processor. Its manifest records F13_RATIFIED_CHAT at 2026-09-29 22:42 MYT and deferred ceremonial sealing. The inspected source still contains earlier drafting labels; this card records the manifest's claim without independently certifying the sovereign receipt or a VAULT999 append.

A candidate fine-tuned model may assess evidence and recommend outcomes. Authority remains delegated by the human sovereign, validated by the kernel, and enforced by actuators. Neural confidence cannot create permission. ARIFOS_ML_FLOORS=1 enables SBERT-backed scoring in the inspected floor auditor, not an i-ARIF/Ollama model switch.

Local world-model ledger: observation, not dataset import

Observed on 2026-10-04, approximately 13:33 MYT:

Reported chain head: 9acbc223ab602787fa8dfa708b43871bcfc147e53ee407883d8ae53d594b425c.
These are values read from local ledger metadata, not independently parsed record counts or verified hash-chain integrity. The 135 eligibility flags do not establish 135 valid training examples, and these local traces have not been imported into this Hub dataset by this documentation update.

The inspected VPS copy of distill_experience.py contains an outcome classifier that treats explicit false outcomes as failures, missing or ambiguous outcomes as UNKNOWN, and marks those cases ineligible. Local file SHA256 at inspection: 8b66f6b8880a4031b66a3501bf0abed2453658cd29f3b333375ce57da9d4440d. Metadata reports eight additional failure records with eligibility remaining 135; failures were retained in the evidence ledger, not deleted. The inspected GitHub main copy still contains the prior implementation. Commit synchronization, independent tests, historical reclassification, and full corpus validity remain unverified.

Readiness limits

The active 1,435-pair corpus count is independently verified against live repository files (1,322 SFT + 96 DPO + 17 VAL) with 100 held-out evaluation pairs. Before training: independently validate typed outcomes and provenance, reconcile ledger integrity, audit legacy eligibility labels, deduplicate, separate evaluation by task/source family, and measure shadow performance. No current CPU latency, training budget, privacy guarantee, or CLM efficiency gain is established by this update.

A CLM continuity pilot requires a host-harness adapter that assembles subsequent model requests from accepted context edits. An editable file or MCP editing tool alone does not provide that behavior. Authority enforcement remains outside the mutable working context.

Citation

If you use this dataset for training or evaluation, please cite:

@dataset{fazil2026iarifcanon,
  author    = {Muhammad Arif bin Fazil},
  title     = {I-ARIF CANON: Governed Execution Substrate Training Corpus},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ariffazil/I-ARIF-CANON},
  license   = {AGPL-3.0},
  note      = {Constitutional training corpus for Project I-ARIF. Companion datasets: ariffazil/AAA through ariffazil/FFF}
}

Sovereign

Muhammad Arif bin Fazil (F13). DITEMPA BUKAN DIBERI.

Federated update — 2026-10-04. Reconciled composite graph and YAML counts against active corpus files (1,435 active + 100 held-out). No dataset data files were mutated; only documentation and metadata were updated. DITEMPA BUKAN DIBERI — Forged, Not Given.

🏛️ Sovereign Model Roadmap — Updated 2026-10-18

Under F13 Sovereign Directive (2026-10-18):

• Canonical Model Role: The I-ARIF model is architected as an 888-APEX Constitutional Judge (1B–3B parameters), not an unconstrained text generator.
• Primary Corpus:
  • i-arif-sft.jsonl: 1,308 supervised constitutional pairs (F1–F13 floor tables, arifOS doctrine, Nusantara synthesis).
  • i-arif-dpo.jsonl: 96 preference pairs enforcing anti-fabrication, ungrounded refusal penalties, and identity integrity.
  • agentic-trajectories-rl.jsonl: OpenAI function-calling format multi-step tool-use trajectories.
• Deployment Substrate: Ephemeral QLoRA fine-tuning (RunPod GPU), exported to GGUF (Q4_K_M / Q8_0), deployed on local KVM CPU (AMD EPYC Zen 4 AVX-512) behind the arifOS LiteLLM Federation Router.
• Auditor Verification: Fully verified against primary repo files; zero modification to immutable data payloads.
• Ditempa Bukan Diberi: Sealed under F13 Sovereign Authority.