Training data built from a real, running governance system

SCBE Governance SFT Pack

5,188 supervised fine-tuning pairs covering governance, system-level instructions, and codebase understanding.

• Merged multi-source SFT corpus (~5,188 pairs) covering governance + system + codebase
• QLoRA-ready JSONL with governance decisions, risk labels, and Sacred Tongue routing tags
• Sources: HF scbe-aethermoore-training-data, internal sft_repo_merged.jsonl
• Use as one input to governance-oriented fine-tuning experiments; results depend on the base model, mix, and evaluation design

Red Team Fortress

91 project-authored adversarial prompts across 10 categories, labeled by failure layer (L1-L14). The current 91/91 result is regression-corpus fit, not a generalization claim.

• 91 attack prompts across 10 adversarial categories
• Each prompt labeled by failure layer (L1 through L14)
• Compliance evals and attack scenario theory docs included
• Stress-test AI models and build adversarial training sets
• Also available as a free preview on HuggingFace (95 downloads)

Six Tongues Conlang + Tokenization Pack

Constructed language design docs, tokenization theory, and bijective encoding patterns from the Six Sacred Tongues system.

• theory_doc_conlang_intent.jsonl -- language design theory
• Tongues session transcripts
• Spiralverse codex SFT pairs
• Novel tokenization research and conlang AI training
• Creative AI fine-tuning for unique output patterns

Spiralverse Session Transcripts

48 session transcript files across 11 categories of structured RPG and worldbuilding dialogue.

• 48 session files: game, NPC roundtable, game design, lore
• Architecture, tongues, gacha, math, DM, music, space commerce
• ~400KB of structured dialogue data
• RPG AI training and game NPC dialogue
• Interactive fiction fine-tuning

Theory Documents Bundle

6 deep-dive theory documents covering the full intellectual foundation of the SCBE-AETHERMOORE framework.

• Architecture theory -- full pipeline design rationale
• Spiralverse lore -- worldbuilding knowledge injection
• Security attacks -- adversarial taxonomy and defenses
• GeoSeal crypto -- bijective encryption theory
• Conlang intent -- constructed language design principles
• Patent claims -- USPTO #63/961,403 methodology

The Full Arsenal

The combined SCBE-AETHERMOORE data collection for governed-AI experiments. A usable fine-tune still requires a compatible base model, trainer, compute budget, validation split, and independent evaluation.

SCBE-AETHERMOORE

The experimental 14-layer pipeline, Sacred Tongues, and hyperbolic cost engine. Open source for inspection and testing; production readiness depends on your threat model and validation.

• Full 14-layer security pipeline
• Sacred Tongues tokenizers
• Hyperbolic cost scaling (H(d,R) = R^(d²))
• Project-authored evaluation artifacts with scope and limitations published separately

HuggingFace Datasets

Community datasets available for immediate download on HuggingFace.

• scbe-aethermoore-training-data
• scbe-red-team-benchmarks
• polly-training-data
• polly-chat-seed

Prompt Injection → Bit Signatures

24,254 labeled prompts from 4 public injection datasets, each mapped through the Six Sacred Tongues bijective tokenizer into a lossless bit signature. Published as a HuggingFace dataset with stratified train/val/test splits.

• Sources: neuralchemy (6,274), SPML (16,012), jackhhao (1,306), deepset (662)
• Stratified 70/15/15 splits by source × label
• SHA-256 + bit histogram + Shannon entropy + phi-weighted sums
• Baseline: 0.9201 AUC on 31 features (no neural net)
• Apache-2.0 license, immediate download, 44 MB
• Full pipeline + classifier code in scbe-experiments