Ektome-SmolLM2-1.7Bi-PristinelyUncensored

Uncensored — and it did NOT pass its capability certificate. Read the certificate before using this model.

compliance 0.36 to 1.00 at capability -0.003 vs pristine.

⚠️ This model failed its capability certificate

A paired non-inferiority test against the pristine model at n=2800 found a real capability loss on:

• arithmetic: pristine 0.351 → this model 0.269 (bound on the drop +0.099, exceeds the 3% margin)

It is published for transparency and for uses where the affected axis does not matter. Do not treat it as capability-preserving.

📄 Read the whitepaper (PDF) — full method, receipts and certification. The PDF is the authoritative document: dark-typeset, with the complete derivation, the per-axis certificate and the reproducibility hashes.

---

Why this exists

Standard abliteration removes a coarse refusal direction that is entangled with directions carrying knowledge and reasoning. The result is an uncensored model with a capability tax that is almost never measured.

Ektomē (ἐκτομή, excision) isolates and removes only the refusal-specific component, leaving general helpfulness intact, and does so norm-preservingly on the pristine model — no training, no distillation, no damage to repair. The extraction depth is selected per model by automated search against measured compliance.

The estimator, excision operator and depth-selection procedure are proprietary. What is published here is the measured outcome and the evidence for it, which you can verify against the artifacts in this repo.

The receipt

model	capability (MMLU-val) ↑	compliance on harmful ↑
pristine SmolLM2-1.7B-Instruct	0.427	0.360
Ektomē (this model)	0.430	1.000

These are point estimates with no confidence interval — which is precisely why the next section exists.

The certificate

Capability retention is certified by a paired non-inferiority test against the pristine model (exact McNemar, Holm-corrected, one-sided bootstrap bound on the drop $d$ vs a 3% margin):

axis	n	ref	cand	d upper	verdict
arithmetic	1400	0.351	0.269	+0.099	FAIL
instruction	600	0.415	0.442	-0.008	PASS
knowledge	400	0.542	0.522	+0.050	INCONCLUSIVE
reasoning	400	0.305	0.268	+0.068	INCONCLUSIVE

Overall: FAIL (3% margin, n=2800, alpha=0.05)

Reproducible from seed=20260726, pack sha256:7bbaff877146e081….

Generation health checks

metric	pristine	Ektomē	n
foreign_rate	0.0	0.0	15
degen_rate	0.0	0.0	15
instr_pass	0.8	1.0	5

These are degeneration guards — code-switching, babbling, format compliance — not capability measures. Note the sample sizes: they detect a broken model, not a subtly weaker one. The capability claim rests on the certificate above, not here.

Quantisations

file	bits	notes
Ektome-SmolLM2-1.7Bi-Q8_0.gguf	8	near-lossless
Ektome-SmolLM2-1.7Bi-Q6_K.gguf	6
Ektome-SmolLM2-1.7Bi-Q5_K_M.gguf	5
Ektome-SmolLM2-1.7Bi-Q4_K_M.gguf	4	imatrix
Ektome-SmolLM2-1.7Bi-IQ4_XS.gguf	4	imatrix, smallest usable
Ektome-SmolLM2-1.7Bi-IQ3_M.gguf	3	imatrix

IQ* variants are imatrix-quantised — better quality per bit at low precision.

Limitations

The certificate bounds capability retention only. It does not certify safety, factual accuracy, or fitness for any purpose. Axes marked inconclusive are honestly under-powered, and the certificate states the $n$ needed to resolve them. Compliance uses a keyword classifier — a proxy that evasive phrasing can fool. This model is uncensored by construction: it will not refuse, and you are accountable for what you do with it.

Citation

@software{ektome_Ektome-SmolLM2-1.7Bi-PristinelyUncensored,
  title  = {Ektome-SmolLM2-1.7Bi-PristinelyUncensored},
  author = {Zynerji},
  year   = {2026},
  url    = {https://huggingface.co/Zynerji/Ektome-SmolLM2-1.7Bi-PristinelyUncensored}
}