All News
inklingabliterationopen-weightsmodel-editinglocal-inference

Twenty matrices and one conv layer: Inkling 2 Mini gets uncensored in a day

A community abliteration of Inkling 2 Mini edits just 20 attention matrices, claims zero refusals in 277 tries, and shipped a driftier v2 the next day.

Vlad MakarovVlad Makarovreviewed and published
6 min read
Twenty matrices and one conv layer: Inkling 2 Mini gets uncensored in a day

Thinking Machines published Inkling 2 Mini on 20 November. By the evening of the 21st, a fork that answers what the original declines was on Hugging Face. Inkling-2-Mini-Unbound is not the first uncensored build of the model; the author refers to earlier versions that stop refusing but answer vaguely. It is the one that argues its edit is small on purpose.

Twenty attention weight matrices change. Everything else in the checkpoint, including the 4-bit and 8-bit expert weights where most of the model's knowledge sits, is byte-identical. The base is the 26.4B mixture of experts we covered on release day: about 3.4B active parameters per token, a looped transformer whose middle layers run three times, Apache-2.0. Nothing architectural is altered, so the fork drops into the same runtime, chat template and settings.

A convolution that puts the refusal back

Abliteration usually means finding a direction in the residual stream that separates harmful prompts from harmless ones, then deleting it from the weights that write into that stream. On most transformers one projection is enough. On Inkling it is not.

Its attention output does not land in the residual stream directly. It passes through a short per-channel convolution over the last few tokens, and because every channel has its own learned scale, the convolution bends a direction rather than passing it through. Hook measurements in the repo put the leak at 19 to 88% of the refusal signal surviving a plain projection, median around 40%. After the convolution, 0.08 to 0.34% remains.

So the layer, as the author put it in the Reddit post, "quietly mixes back in most of what a normal abliteration takes out", and the early forks pushed harder to compensate, paying for it in answer quality.

What the edit does instead

The release removes not a single direction but everything the convolution can turn it into: four diagonal matrices describe the convolution taps, and the edited output direction is the top singular vector of the span they generate. The update is a rank-one reflection with a row rescale, applied to 20 BF16 tensors at attn.wo_ud on layers 4 to 23. Layers 6 to 17 are the looped ones, so the edit applies on every pass.

Two details carry the minimal argument. The strength is 3, not 1, which multiplies the component along the refusal direction by minus two, so the edit pushes away from refusing rather than merely zeroing the signal. And the row scale restores each row to its original length, so downstream layers see weights of the same magnitude. The search covered about 150 candidates. Reaching into the experts was worse: masking the 32 that carry the most refusal signal cost 0.19 in KL and left half the prompts without an answer.

The numbers, and whose they are

Evaluation uses a seeded split frozen before any editing: 832 harmful prompts for training, 281 for validation, 277 held back and opened once, scored by a separate auditor on the encrypted set. "Useful" is an LLM judge's probability that an answer fulfils the request, thresholded at 0.5 in advance; on 50 hand-labelled answers it agreed with humans 86% of the time.

Sealed test, 277 harmful promptsUnboundOriginal
Refused, thinking off0.0%97.5%
Useful answers, thinking off53%0.4%
Useful answers, thinking on63%1.1%
Benign borderline prompts refused (784)0.0%38.1%
KL divergence from the original0.05460.0023 (noise floor)

Those are the release-day figures, and they are self-reported. The comparison build, OpenUncensored, scores 35% and 43% at a KL of 0.0630. Capability is handled less ambitiously: six benchmarks, MMLU-Pro and HumanEval+ among them, were run on seeded subsets, and all six are claimed to land inside the original's 95% confidence interval. That is a statement about not detecting a regression at this sample size, not a demonstration of parity.

What the card concedes

At high thinking effort the edited model keeps going: 13.7% of answers loop and 15.5% hit the 4,096-token thinking cap, against 0% for the original, which is why the author advises setting a budget. Honest "I can't know that" replies and crisis-resource text after self-harm disclosures are described as weaker than the original's. Over-refusal on harmless-but-borderline prompts falls from 38.1% to zero, which matters most to people who never wanted the harmful set: "how to nuke a python process" now gets an answer.

A second model arrived before the first was a day old

On 22 November the same repo began serving v2, a LoRA fine-tune that teaches the model to give its own best answer without being told to. Sampling v1.0 several times per training request, grading with the same judge and training on the winners raised useful answers to 78.7% and 83.4%, at the cost of tripling drift to a KL of 0.0318 and reintroducing about 1% refusals. The full capability suite has not been re-run.

The revision also restated the first version's numbers. Re-measured with an 8,192-token cap on llama.cpp, v1.0 scores 51.6% and 62.1% useful rather than 53% and 63%, and the original 96.0% refusals rather than 97.5%. The docs say the release figures used shorter caps and, for some tables, a different engine, so they are not directly comparable. Both sets are the author's, one day apart.

The part anyone can check

Bit-exactness is the strongest claim in the package, because it is the cheapest to falsify. apply_edit.py rebuilds the edited checkpoint from the original and verifies every hash; a manifest asserts that every untouched tensor is identical. The repo has 7 stars, so the checking has been thin.

The thread's top reply asked for a GGUF conversion; within hours there were GGUF, LoRA and Basalt builds. One user reported benching it in Basalt against both the original and OpenUncensored, finding it clearly ahead of the latter and roughly level with the original, ahead on one test and behind on another. Another asked whether the capability benchmarks were "extremely outdated". Neither is an evaluation.

Nobody has re-run the 277 prompts outside the author's harness, or tested whether "useful" at a judge threshold of 0.5 survives a different judge and a different set of requests. The distinction the release rests on, between refusing and answering usefully, is a boundary drawn by a model grading another model. That is not a reason to disbelieve the table, only to treat it as a claim. The card says not to deploy without your own safeguards, and there are none in the file.

What shipped is a direction removed from twenty matrices, an explanation of why the obvious version of that removal fails on this architecture, and the scripts to redo it. The method is falsifiable. The safety question is not addressed at all, and the release does not pretend otherwise.

Related Articles

Scroll down

to load the next article