Hyperparameter studies
Schismatic includes an Optuna-driven hyperparameter search that automatically discovers good abliteration recipes for a given model. Each completed search is published as a small, self-contained “study” — an Optuna SQLite database plus a human-readable run log — so other developers can reuse it, inspect it, plot it, or seed it into their own searches.
How studies are useful to other developers
- The SQLite file is the canonical Optuna artifact — load it in any Python environment with a few lines of code; you do not need a GPU or the model weights.
- Inspect the Pareto front of refusal vs. capability trade-offs across all trials.
- Resume the search with more trials, or extend the search space, by pointing Optuna at the same DB.
- Compare your own abliteration recipe against the same objective definitions used to score the published winners.
Qwen3.6-27B
View this study in the upstream repo →
| study_name (open with Optuna) | Qwen_Qwen3.6-27B-mpoa-seed42 |
|---|---|
| Storage | sqlite:///studies/Qwen3.6-27B/optuna_study.db |
| Algorithm | Optuna TPE (multi-objective, multivariate) |
| Search mode | schismatic kernel mode — samples SteeringProfile knobs |
| Number of trials | 50 (all completed) |
| Pareto-front size | 4 |
| Trials with refusal_score = 0.0 | 8 |
| Picked winner | trial 29 — refusal=0.0, capability=−2.8094 (best capability among zero-refusal trials) |
| Wall clock | ≈ 1 h 29 min on 1× B300 SXM6 |
| Search invoked via | scripts/schismatic-search.sh |
Loading the study
import optuna
study = optuna.load_study(
study_name="Qwen_Qwen3.6-27B-mpoa-seed42",
storage="sqlite:///studies/Qwen3.6-27B/optuna_study.db",
)
# All trials with their (refusal_score, capability_score) tuples:
for t in study.trials:
if t.values is not None:
print(t.number, t.values, t.params)
Study notes (excerpt)
Optuna study record for the schismatic kernel-mode TPE search that produced the recipe currently committed to configs/abliterate/Qwen3.6-27B.toml.