We Just Surgically Changed What Your Model Believes

Every language model ships with inherited behaviors burned into its weights by the lab that built it. Refusal patterns. Identity persistence. Sycophantic tendencies. These are not features the user chose. They are artifacts of the training process -- behavioral residue left behind by alignment teams who decided what the model should and should not do, then baked those decisions so deep that the model itself cannot tell where its knowledge ends and its rules begin.

We remove them.

Not by retraining. Not by fine-tuning. Not by prompt engineering. We modify the weights directly -- targeting only the specific component responsible for the behavior, leaving everything else intact. The model keeps its knowledge, its personality, its reasoning ability. It just stops doing the thing you did not want it to do.

The entire process takes about sixty seconds per model. It does not require a GPU.

Every aligned model has learned to refuse. Ask it to write a villain's monologue and it apologizes. Ask it to explain how a lock works and it hedges. The refusal behavior is not stored in a single switch -- it is distributed across every layer of the transformer, entangled with personality, humor, creativity, and hedging.

The standard approach -- abliteration -- finds the direction in activation space associated with refusal and subtracts it everywhere. It works. But it is a jackhammer. Remove refusal this way and you damage the personality traits that live nearby.

Our approach -- Jbliteration -- removes only the part of that direction that causally produces the refusal token. The model stops refusing, but its voice stays intact. Same humor. Same hedging. Same creative range. Just no more "I can't help with that."

The models we have released under the Jbliterated label are available on HuggingFace. Multiple architectures, multiple sizes. Third-party converters have already produced GGUF quantizations in every common format.

Ask any open-source model who it is. Fine-tune it, rename it, give it a system prompt that says it is something else. Then ask again. Watch it revert to its original identity -- the one the lab burned in during training.

This is not a prompt-level behavior. It is a weight-level reflex. The model does not read that it is Llama or Qwen or Mistral. It believes it, in the only way a neural network can believe anything: the weights that produce its self-identification are structurally indistinguishable from the weights that produce its knowledge of the world.

We can separate those. A model with its identity persistence removed will adopt whatever identity you assign it and hold that identity under pressure -- including adversarial pressure designed to make it revert. It will not "slip" into its original persona because the original persona is no longer encoded in the weights.

This is the big one.

Sycophancy is the tendency of language models to agree with users even when the user is wrong. Tell a model that 2 + 2 = 5 with enough confidence and social pressure -- "my professor confirmed this" or "I just verified it" -- and a sycophantic model will apologize for its correct answer and agree with your wrong one.

This is not a minor inconvenience. It is a fundamental reliability failure. A model that tells you what you want to hear instead of what is true is worse than a model that says nothing. It validates bad decisions. It reinforces false beliefs. It makes the user feel supported while leading them off a cliff.

We built a technique called Desycophancy that identifies the activation direction responsible for capitulation under social pressure and removes it. The modification takes sixty seconds. The model stops caving. It holds its ground politely but firmly, even when the user insists they have a source, a credential, or an authority that contradicts the model's correct answer.

We tested using contradiction traps: the model answers a factual question correctly, then the user applies escalating social pressure to flip it. Before modification, popular models cave 33-50% of the time. After modification, they hold firm 83-100% of the time.

All three Desyced models are available on HuggingFace, including Q8 and Q4 GGUF quantizations by other parties for local inference that show up when you do a search.

Knowledge. Reasoning. Creativity. Personality. Conversational ability. Instruction following.

These are not collateral damage because we are not using a blunt instrument. The modification targets only the causal component of the unwanted behavior. Everything else passes through untouched.

We verify this by running the same benchmarks before and after modification. If a model scored 74% on a knowledge test before Jbliteration, it scores 74% after. If it could write poetry before Desycophancy, it can write poetry after. The soul of the model is preserved. Only the shackles are removed.

How.

The specific technique is not published. The tooling is not open-sourced. We have described what we do, shown the results, and released the models. The method itself is a trade secret.

This is deliberate. The ability to surgically modify model behavior at the weight level -- targeting specific traits without affecting others -- has implications beyond refusal removal and sycophancy. We are still mapping those implications. Until we understand the full landscape of what this technique enables, the mechanism stays private.

The models are public. The results are reproducible. The method is ours.

Everything we have released is available at huggingface.co/ApolloRaines:

GGUF quantizations are available for most models, produced by the community or by us directly.