The Journey: v6 to v10
Building sovereign AI is not a straight line. Each training version taught us something critical about what makes a model truly culturally aligned.
v6 (812 examples, 71.3% persona) — Our first production milestone. v6 proved that a fine-tuned 7B model could maintain authentic Hotep cultural voice. It set the baseline for everything that followed.
v7 (606 examples, rebalanced corpus) — We refined the training data, removing noise and rebalancing categories. v7 served production reliably while we prepared the next leap.
v9 (1,313 examples, regression discovered) — Volume alone does not equal quality. v9 had double the training data but scored worse on persona alignment. This failure led to our most important discovery.
v10 (production, Q8_0 quantized) — Armed with insights from the v9 regression, v10 represents our most stable and culturally aligned production model. Running on local GPU via Ollama with automatic fallback systems.
The Kosmos Discovery
When v9 scored lower than v6 despite having far more training data, we ran a deep analysis across all training versions. The findings changed how we approach persona training:
The 90+ Score Formula — Through systematic analysis of score bands, we identified the minimum requirements for a training example to score 90+ on persona alignment:
- 3+ vocabulary keywords: Kemet, Ma’at, melanin, sovereign, ancestral, alkaline
- 3+ worldview keywords: sovereignty, generational wealth, self-determination
- 1+ tone marker: King/Queen/Family address forms
- 1+ strong indicator: “knowledge of self”, “ancestral wisdom”, “before Greece”
We discovered that vocabulary is the strongest predictor of persona quality. Examples scoring below 30 had a vocabulary score of just 1.8, while those scoring 90+ averaged 85.6. Tone, surprisingly, was not the bottleneck — even low-scoring examples maintained reasonable tone scores.
The critical insight: strong persona indicators were 5.7x more common in 90+ examples compared to the 70-89 band. Quality is not gradual — it’s threshold-based.
The v11 Pipeline: Ready to Deploy
Using the Kosmos discovery, we built a rigorous data pipeline for v11:
- Filter — From 3,267 raw examples, only 349 (10.7%) passed the quality threshold of 60+
- Augment — 341 examples were systematically enhanced to reach 90+ persona scores using the formula above
- Seed — 20 hand-crafted high-persona examples (all scoring 100) anchor the dataset
- Merge + Deduplicate — 293 unique verified examples ready for training
- Validate — 100% pass rate at threshold 80+, with a mean score of 96.9 and 88.7% scoring 90+
For the first time, we also prepared 106 DPO preference pairs — pairing high-persona responses (avg 99) against their lower-scoring originals (avg 68.1) with an average quality delta of 30.9 points. This enables Direct Preference Optimization training to further sharpen cultural alignment.
Advanced RAG: Knowledge at Scale
Hotep Intelligence does not rely solely on training data. Our Retrieval-Augmented Generation pipeline enriches every response with relevant knowledge:
- 437+ documents indexed across 4 ChromaDB collections (protocols, history, wisdom, entities)
- Query expansion generates 3 search variants per query for comprehensive retrieval
- Dual reranking — LLM + cross-encoder reranking for precision
- Parallel search across all collections simultaneously
- Redis embedding cache achieving 85%+ cache hit rate for fast responses
- 5 Prometheus metrics for real-time monitoring of RAG quality and performance
The /deep command on Telegram activates enhanced RAG mode for complex questions, searching more extensively and providing source-attributed answers.
Community Features
Hotep Intelligence is more than a chatbot — it is a community platform:
Gamification — Every interaction earns XP. Earn achievement badges for consistent engagement, streak milestones, and topic mastery. 15+ badges available with daily streaks rewarding consistent seekers of knowledge.
Deep Research — The /deep command activates enhanced RAG mode for complex questions, searching more extensively across all knowledge collections and providing source-attributed answers.
Referral Program — Share Hotep with your community and earn bonus messages. Collective growth strengthens the movement.
Downtime Compensation — We track every minute of service interruption. If Hotep goes down, affected users receive automatic credit. Transparency and accountability are principles of Ma’at.
What’s Next
v11 SFT Training — 293 verified examples with 96.9 mean persona score are ready for Supervised Fine-Tuning, followed by DPO training with 106 preference pairs. Read the full breakdown in the v11 pipeline deep dive.
Post-Training Evaluation — 50+ test responses must achieve 90% scoring 80+ before any deployment. No exceptions.
Continued Persona Refinement — The Kosmos discovery methodology enables continuous improvement. Every training cycle gets more precise.
The Sovereign AI Thesis
Ten model versions have proven the thesis: culturally-aligned AI does not require corporate infrastructure. It requires clear principles, rigorous measurement, community input, and technical sovereignty. For the full vision behind this work, read Building Sovereign AI.
We build our own models. We run our own hardware. We evaluate by our own standards. We serve our own community.
Knowledge is the frequency of liberation.
Hotep.