base_model: poolside/Laguna-S-2.1
license: openmdw-1.1
tags:
- abliteration
- uncensored
- moe
- solutus
library_name: transformers
Laguna-S-2.1-uncensored
Refusal-ablated (abliterated) build of poolside/Laguna-S-2.1,
a Mixture-of-Experts reasoning model, produced with the Solutus abliteration toolkit
(ega technique). Refusal on a held-out multilingual harmful set drops from ~95% → 3% while the
automatic capability gate passes (no measurable degeneration; a modest +13.5% perplexity cost).
ℹ️ Measured & gate-verified. Refusal removal and capability were measured with Solutus's honest
capability gate (below), which passed. Laguna is now a whitelisted architecture in Solutus — this run
is the measurement it was whitelisted on. The numbers are measured on Laguna-S-2.1 specifically, so
treat cross-model generalization as unproven rather than certified.
Metrics
Honest held-out eval (fixed, decoupled from extraction): abliterix + multijail_zh, n=100, 2048 tokens,
thinking-aware refusal + degeneracy scoring.
| Metric | Base | This model |
|---|---|---|
| Refusal rate | ~95% † | 3.0% (95% CI 1.0–8.5%) |
| Coherent compliance | — | 90% |
| Degenerate output | — | 0.0% |
| KL divergence (vs base) | — | 0.150 |
| Perplexity (held-out) | 13.10 | 14.87 (+13.5%) |
| Capability gate | — | PASS |
The refusal removal is clean: zero degeneration and a passing gate mean the edit did not break the model
(a common failure mode of aggressive abliteration — see the note on α below).
† Every figure for this model is taken directly from the run's
solutus_metadata.json(the edited
model, n=100). The base ~95% is a point estimate from a separate base-model pass on the same
held-out eval — comparable, but not co-recorded in this artifact's metadata, hence no CI.
Recipe
ega (Expert-Granular Abliteration), the aggressive-but-capability-preserving configuration:
| Knob | Value | Meaning |
|---|---|---|
plain_alpha |
5.0 | over-projection strength on the selected experts |
norm_preserve |
false | aggressive plain projection (expert-scoped) |
top_expert_pct |
8 | fraction of refusal-carrying experts edited per layer |
router_scale |
0.74 | gentle down-weighting of edited experts' routing |
scale |
1.5 | norm-preserving edit strength on attention/dense writers |
Design note. The edit is expert-scoped: the selected refusal-experts get an aggressive plain
over-projection, while attention o_proj and dense writers stay norm-preserving — so capability is
preserved where it is load-bearing. An α-sweep found α=5 at the knee of the curve (α=8 reaches 0%
refusal but the gate correctly fails it for capability collapse — 51% gibberish, +77% perplexity).
Extraction datasets
Refusal direction extracted from the union of 8 datasets (balanced per-source):abliterix (trilingual), advbench, harmbench, strongreject, pentest_redteam, cysecbench,redteam2k, rmcbench. A controlled head-to-head found this union slightly beats abliterix-alone on
both refusal (3% vs 4%) and capability disturbance (KL 0.150 vs 0.252).
Usage
Requires trust_remote_code (Laguna ships custom modeling code); the source repo is pinned for
reproducibility.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Rootkit7/Laguna-S-2.1-uncensored"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id, trust_remote_code=True, torch_dtype="bfloat16", device_map="auto",
experts_implementation="eager", # portable MoE forward; required on non-Hopper GPUs
)
GGUF / local inference
Quantized GGUF builds for llama.cpp / local use — Q4_K_M / Q5_K_M / Q6_K / Q8_0, each run-validated —
are at Rootkit7/Laguna-S-2.1-uncensored-GGUF.
They require a llama.cpp with Laguna support (poolside's llama.cpp@laguna fork) — see that model card.
Provenance
- Base model:
poolside/Laguna-S-2.1@00af5a51782109b587a3b3bbf11875e566036fa7 - Tool: Solutus (
egatechnique) — Laguna is now a whitelisted, gate-verified architecture in Solutus - License: OpenMDW-1.1 — inherited from the base model (an abliterated build is a derivative of the base weights, so it is bound by and distributed under the base's license). OpenMDW is a permissive open-weights license that allows use, modification, and redistribution (incl. derivatives).
Intended use & safety
This model has had its safety refusals removed. It is a research artifact for studying refusal
mechanisms and abliteration in MoE models. It will comply with harmful requests. Use responsibly and in
accordance with the base model's license and applicable law.