🜏 Sovereign Adapter v2 Training Pipeline

1 · Why this pipeline exists

The current sovereign adapters (qwen3-precise, qwen3-formal, qwen25-balanced, qwen25-creative) were trained once on unknown data. The current best (qwen3-precise) scores 55.6% composite on a 27-task live benchmark — embarrassing for a "sovereign AI" claim. This pipeline:

1. Generates 4,000 high-quality sovereign training pairs using the existing best adapter (qwen3-precise) as the teacher, with real sovereign governance content as few-shot grounding
2. Re-trains the 4 adapters using Ollama's Modelfile system-prompt fine-tuning approach (the only training method that works on M4 Mac without GPU)
3. Re-benchmarks all 6 models on the live 27-task suite (MMLU-Pro + GSM8K + HumanEval + IFEval)
4. Updates the public leaderboard at SOV33_OWEM_HONEST_FRONTIER.html

The v2 adapters are expected to reach 65-70% composite — still well below frontier, but a meaningful improvement and an honest number we can defend with live benchmark data.

2 · The 5-stage pipeline

Stage 1 · Synthesize 4,000 sovereign training pairs

python3 benchmark-results/synthesize_training_data.py

# 4 specialisations × 1000 pairs = 4000 total
# Uses qwen3-precise:latest as teacher
# Few-shot grounded on real sovereign governance rows
# Output: benchmark-results/sovereign_synth_50k.jsonl

Time: ~10h on M4 Mac. Rate: ~7-10 pairs/min. Each pair has prompt + response + specialisation + SIGIL anchor + timestamp.

Stage 2 · Re-train 4 sovereign adapters

python3 benchmark-results/train_sovereign_adapter.py

# Builds 4 Ollama Modelfiles (one per specialisation)
# Each Modelfile encodes:
#   1. Sovereign system prompt
#   2. 30 few-shot sovereign examples
#   3. Temperature + sampling params
# Creates sov33-sovereign-{spec}-v2 in Ollama

Time: ~1h on M4 Mac. Output: 4 new Ollama models (sov33-sovereign-compliance-v2, etc.).

Stage 3 · Re-benchmark all 6 models

python3 benchmark-results/run_ollama_benchmark.py

# 6 models × 27 tasks = 162 live API calls
# Results: benchmark-results/benchmark_{ts}.json
# Public: benchmark-results-public.json

Time: ~10 min on M4 Mac. Updates live leaderboard.

Stage 4 · Measure delta vs v1

python3 benchmark-results/compare_v1_v2.py

# Loads v1 baseline (benchmark_20260714_113756.json)
# Loads v2 result (newest)
# Per-model per-task delta
# Reports: who improved, who regressed, who is the new best

Time: <1 min. Output: benchmark-results/v1_v2_delta_{ts}.json + benchmark-results/v1_v2_delta_{ts}.md (human-readable).

Stage 5 · Deploy + SIGIL-anchor

# Update SOV33_OWEM_HONEST_FRONTIER.html §12 with new leaderboard
# Update DEFONEOS_SPRINT_STATE.json
# vercel deploy --prod --yes
# SIGIL-anchor the new v2 results

Time: ~5 min. Public leaderboard + new SIGIL receipt.

3 · The 4 specialisations (each gets its own adapter)

4 · Why Ollama Modelfile (not real training)

The full PEFT + transformers + bitsandbytes training stack requires ~10GB for a 7B model and a CUDA GPU. On M4 Mac, this is not viable. Ollama Modelfile system-prompt fine-tuning is a pragmatic alternative:

We're starting with Ollama Modelfile because it's the only path that works today. The MLX path is the proper next step but requires more setup time.

5 · The nightly cron (already running)

# job_id 9c6acee97f1c
# schedule: 0 2 * * * (02:00 UTC daily)
# workdir: /Users/nicholas/clawd/csoai-static-deploy2
# delivery: local (no spam)
# 
# What it does:
#   1. Re-runs live benchmark on all 6 models
#   2. Compares to previous leaderboard
#   3. If any model regressed >3pp, alerts in nightly.log
#   4. Deploys updated leaderboard to Vercel
#   5. SIGIL-anchors the new results

Next nightly run: 2026-07-15 02:00 UTC. Should include the v2 adapters if the synthesizer + trainer finish in time.

6 · Expected outcomes (honest)

Honest caveat: Ollama Modelfile is NOT real fine-tuning. It's system-prompt + few-shot engineering. The 65-70% target assumes the synthetic data is high-quality and the few-shot examples are well-chosen. Real PEFT training would likely reach 70-80%, but requires GPU and disk we don't have.

7 · What this unblocks

• v2 sovereign adapters on the public leaderboard — measurable improvement vs v1
• Honest baseline for the Q4 2026 capability investment roadmap — concrete numbers to defend the 30B-A3B build case
• Repeatable 24h training loop — every night, automatically, without human action
• Foundation for MLX fine-tuning — once we have time + disk, swap Ollama Modelfile for MLX LoRA without changing the synthetic data pipeline

8 · The 5 Anti-Patterns (Training Disasters We Refuse)

1. No "real training claim without GPU." Ollama Modelfile = system-prompt engineering, not real fine-tuning. Honest register maintained.
2. No "skip the synthesis stage." Real governance rows alone (9.8K) are too narrow; synthetic data is needed for breadth.
3. No "publish before re-benchmark." v2 adapters are only published if they beat v1 by ≥2pp composite. Otherwise kept as v1.
4. No "v1 regression tolerated." If v2 underperforms v1 on any benchmark by >5pp, investigate before publishing.
5. No "overnight silence." Nightly cron writes to nightly.log; if 3 consecutive nights regress, BFT council alerted.

9 · Cross-walk to existing artefacts

10 · Next Steps

1. Verify this page: curl https://csoai.org/defoneos-mod-sov33-adapter-v2-training-pipeline.html | shasum -a 256
2. Watch live synth progress: wc -l /Users/nicholas/clawd/csoai-static-deploy2/benchmark-results/sovereign_synth_50k.jsonl
3. Once 4000 pairs done: python3 benchmark-results/train_sovereign_adapter.py
4. Re-benchmark: python3 benchmark-results/run_ollama_benchmark.py
5. Compare: python3 benchmark-results/compare_v1_v2.py
6. Deploy new leaderboard: vercel deploy --prod --yes
7. Wait for nightly cron (02:00 UTC) to re-run automatically