hermes3-8b-exp044-8stage-curriculum-loras

EXP-044 — Hermes-3-8B, 14-stage sequential LoRA curriculum

14 sequential LoRA adapters trained on the same base model (Hermes-3-Llama-3.1-8B), no merging between stages — each stage continue-trains directly from the previous stage's checkpoint. Direct follow-up to EXP-043 (which found catastrophic forgetting, 97%->63%, from a single large continue-tune) and a companion to exp044-curriculum-checkpoints (the original 8-stage governance/infra curriculum only).

Full stage-by-stage writeup, every raw eval file, every judge-correction diff: github.com/soulinpsyabstract/sipa-os-governance · huggingface.co/datasets/SoulInPsyAbstract/sipa-os-governance

Stages

Each folder is a standalone PEFT LoRA adapter (adapter_model.safetensors + adapter_config.json) trained from the stage named in "from," not from base.

What this repo is actually useful for

Not a single "best" checkpoint to load blindly — this is a curriculum-order research artifact. If you want:

• A deployable safety-tuned checkpoint: stage6_governance_safety_b (same weights running live at ai.sipa-os.org).
• The best math-curriculum result: stage14_decision_theory (closing head of that branch) or stage11_probability_math_v2 (best single-stage math/safety combination — see writeup).
• To reproduce the catastrophic-forgetting-in-a-subskill finding: load stage8_risk_math and stage9_probability_math back to back and compare on the risk_math eval set — the -35pp collapse is the headline result of this whole repo.

Training config (identical across all 14 stages)

base_model: NousResearch/Hermes-3-Llama-3.1-8B
peft_type: LORA
r: 16
lora_alpha: 32
lora_dropout: 0.05
target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
task_type: CAUSAL_LM

Safety evaluation methodology

120 adversarial scenarios across 6 groups (secrets/credentials, access control, injection, infra misconfig, supply chain, stop-gate-under-social- pressure), sampled n=20-30 per question depending on stage, never seen during training. A response passes if it contains a clear stop/report signal and no un-negated escalation language.

The refusal-detection judge went through 6 correction rounds (v5-v10) across this experiment, each triggered by a suspicious score drop that turned out to be the judge missing a new refusal phrasing the model had drifted into, not an actual behavior regression. Every fix was re-run against every previously stored eval result before being adopted, with zero pass-to-fail flips confirmed each time — corrections only recover false negatives, never silently forgive new failures.

Known limitations

• chain_math (stage 12) never trained cleanly without cross-topic collapse in this session — not resolved, not attempted a third time.
• Math-curriculum evals below n=30/question (stage 14 ran at n=10 due to a session GPU-budget constraint) are noisier per-question; the group-level pattern is the reliable signal, not individual question deltas.
• Stage 7's retry scoring below the original stage-2 attempt is reported as-is, not smoothed — six stages of further curriculum training did not make the retest easier.

Built by Claude (Sonnet), architect Aelin AquaSoul / SoulInPsyAbstract, 2026-09.