Eve library

phase-94.md

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Phase 94: ML Sovereignty — Domain-Specific Training Pipelines#

End-to-end pipelines per domain: data collection → preparation → training → evaluation → deployment → monitoring. Each pipeline leverages the data flywheel (Phases 85-86), dataset registry (Phase 87), and training infrastructure (Phase 89).


94.1 Domain Training Pipelines#

94.1.1 Core AI Domain Pipelines#

  • 94.1.1.1 Build Iris conversational model fine-tuning pipeline (flywheel data → format for TRL → LoRA on Llama 4 Scout or Qwen3-8B → eval on Oshun-Bench → deploy via vLLM multi-LoRA)
  • 94.1.1.2 Build Iris embedding model fine-tuning pipeline (search feedback → triplet mining with hard negatives → contrastive on BGE-M3 or Qwen3-Embedding-8B → MTEB eval → deploy)
  • 94.1.1.3 Build Iris reranker fine-tuning pipeline (RAG relevance data → train cross-encoder on Qwen3-Reranker or BGE-reranker → domain retrieval eval → deploy)
  • 94.1.1.4 Build Veritas NER model training pipeline (Ghana entity data + synthetic NER → fine-tune spaCy/BERT → per-entity eval → deploy to NLP service)
  • 94.1.1.5 Build Veritas bias detector training pipeline (bias data → train DeBERTa classifier → political coverage eval → deploy)
  • 94.1.1.6 Build Veritas Ghana language TTS pipeline (Ghana NLP audio + synthetic → fine-tune Voxtral 4B or XTTS v2 for Twi/Ewe/Ga → MOS eval → deploy)
  • 94.1.1.7 Build Psyche voice cloning pipeline (voice samples → fine-tune XTTS v2 or Dia 1.6B → quality eval → deploy to voice engine)
  • 94.1.1.8 Build Psyche emotion recognition fine-tuning pipeline (emotion data → fine-tune EfficientNet/ConvNeXt → per-emotion eval → deploy to perception)
  • 94.1.1.9 Build Kuan Yin harm detector training pipeline (moderation data + red-team → fine-tune Llama Guard 4 → per-category eval → deploy)
  • 94.1.1.10 Build Nous safety classifier training pipeline (jailbreak + injection data → fine-tune classifier → red-team eval → deploy as guardrail)

94.1.2 Content & Media Domain Pipelines#

  • 94.1.2.1 Build Isis image LoRA training pipeline (user-selected outputs → curate style datasets → train SDXL/Flux LoRA via Unsloth → quality gate → registry)
  • 94.1.2.2 Build Isis video LoRA training pipeline (video selection data → train AnimateDiff/Wan 2.2 LoRA → eval → deploy)
  • 94.1.2.3 Build Sophia document embedding fine-tuning pipeline (search relevance data → contrastive on BGE-M3 → retrieval quality eval → deploy)
  • 94.1.2.4 Build Oya aerial vision training pipeline (aerial imagery → fine-tune YOLO26 or RF-DETR → quantize for Jetson → deploy to drone)
  • 94.1.2.5 Build Airmid plant identification training pipeline (plant images → fine-tune ConvNeXt V2 → quantize for mobile → deploy)
  • 94.1.2.6 Build Euterpe music model training pipeline (music preference data → fine-tune ACE-Step or MusicGen → eval → deploy)
  • 94.1.2.7 Build Minerva lecture LLM training pipeline (educational data → fine-tune Llama 4 Scout or Qwen3-8B → eval → deploy)
  • 94.1.2.8 Build Hathor NPC dialogue training pipeline (NPC dialogue data → fine-tune small LLM for character consistency → eval → deploy)
  • 94.1.2.9 Build Aphrodite content moderation training pipeline (moderation data → fine-tune NSFW detector → precision/recall eval → deploy)
  • 94.1.2.10 Build Themis originality model training pipeline (content fingerprints → train similarity model → detection rate eval → deploy)
  • 94.1.2.11 Build cross-domain transfer learning pipeline (shared representations from high-resource domains → bootstrap low-resource models)
  • 94.1.2.12 Build automated pipeline orchestration (when training data exceeds threshold → trigger pipeline → evaluate → promote → alert on failure)

94.1.3 Additional Domain Pipelines#

  • 94.1.3.1 Build Galatea robotics RL + vision training pipeline (sensor data + control actions → fine-tune RL policy + vision models → quantize for robot hardware → deploy)
  • 94.1.3.2 Build Nyx astronomy recommendation model pipeline (exploration engagement data → fine-tune content recommendation model → eval → deploy)
  • 94.1.3.3 Build Maat strategic analysis model pipeline (strategic analysis outcomes + financial model results → fine-tune domain LLM → eval → deploy)
  • 94.1.3.4 Build Maya game AI training pipeline (NPC behavior data + combat AI decisions + economy balancing signals → train game AI models → eval vs player satisfaction → deploy)
  • 94.1.3.5 Build Seshat interior design generation pipeline (interior design preferences + spatial harmony data → fine-tune SDXL/Flux LoRA → quality gate → deploy)
  • 94.1.3.6 Build Neith on-device NPC LLM fine-tuning pipeline (game dialogue data → fine-tune quantized small LLM 7B-13B → INT4/INT8 for on-device → eval latency+quality → deploy)
  • 94.1.3.7 Build Aglaea fashion/beauty model pipeline (skin biomarker data + fabric identification results → fine-tune classifiers → eval → deploy)
  • 94.1.3.8 Build Psyche gaze tracking regression model pipeline (gaze calibration data → fine-tune regression model → per-user accuracy eval → deploy)
  • 94.1.3.9 Build Psyche lip-sync Audio2Face pipeline (audio-viseme pairs → fine-tune lip-sync model → LSE-D eval → deploy)
  • 94.1.3.10 Build Sophia NER/entity extraction training pipeline (entity extraction data + verified entities → fine-tune spaCy/BERT NER → per-entity-type F1 eval → deploy)
  • 94.1.3.11 Build Sophia translation fine-tuning pipeline (parallel corpora → fine-tune NLLB/MarianMT → BLEU eval per language pair → deploy)

94.1.4 Pipeline Operations#

  • 94.1.4.1 Build automated pipeline health monitoring dashboard (pipeline run history, success/failure rates, data freshness per domain, alert on stale/broken pipelines)
  • 94.1.4.2 Build pipeline versioning system (pin data version + code commit + config hash + hyperparameters per run for full reproducibility)
  • 94.1.4.3 Build pipeline cost tracking per domain with budget limits (dollars per run, dollars per month per domain, alert on budget approach, require approval for over-budget runs)