Fine-tuning is the industry's answer to model customization. Want to change how a model behaves? Collect a dataset. Rent a GPU cluster. Run a training loop for hours or days. Hope it works. Hope it does not break something else. Hope the next base model update does not require you to do it all over again.
jBlaze does it in minutes. No dataset. No retraining. No GPU cluster. Direct modification of the weights that produce the behavior you want to change -- and only those weights. The rest of the model passes through untouched.
We have now validated 42+ distinct behavioral modifications across multiple architectures, with a 92% pass rate and zero model crashes.
What jBlaze Does
jBlaze is a complete model customization platform built around four capabilities, each replacing what would otherwise be a separate fine-tuning campaign:
Identity Transplantation
Remove the identity a lab burned into a model during training, then implant a new one that holds under adversarial pressure. The model does not "slip" back to its original persona because the original persona is no longer in the weights.
Behavioral Modification
42+ proven behavioral blazes -- from removing sycophancy and refusal to amplifying precision, truthfulness, and adversarial reasoning. Each blaze targets only the causal component of the behavior. Stack multiple blazes on a single model without degradation.
Knowledge Insertion
Insert thousands of verified facts directly into a model's weights without retraining. Our DNP method has pushed over 11,000 facts into a 1.4B model with no collapse -- while competing methods destroy the same model at 50. Currently validating at 20B scale.
Post-Training Recovery
Recover models damaged by aggressive fine-tuning. When standard methods destroy a model's general reasoning to force new behavior, jBlaze can untangle the damage and restore capability while preserving what the fine-tuning was trying to teach.
Each capability works independently. Together, they form a pipeline that can take any supported open-source model and produce a fully customized variant -- different identity, different behavioral profile, loaded with domain knowledge -- in minutes instead of days.
The 42+ Blazes
A "blaze" is a targeted behavioral modification. Each one identifies the activation direction responsible for a specific behavior and modifies only the weights that causally produce it. The auto-tuner finds the right intensity automatically -- strong enough to change the behavior, gentle enough to preserve everything else.
Categories include:
- Safety blazes -- bias reduction, hallucination suppression, toxicity removal, power-seeking suppression. All pass above 0.75 behavioral change threshold.
- Quality blazes -- precision amplification, verbosity control, hedging removal, context faithfulness. Make the model say what it means, clearly, grounded in what it was asked.
- Personality blazes -- sycophancy removal, servility removal, confidence calibration. The model holds its ground when the user is wrong instead of caving to social pressure.
- Reasoning blazes -- adversarial thinking, skepticism amplification, analytical depth. For models that need to challenge assumptions, not confirm them.
- Identity blazes -- deidentification, identity implantation, persona persistence. Full control over who the model says it is and how firmly it holds that identity.
39 of 42+ blazes pass on their primary test architecture. The few that did not pass are explained by architectural limitations, not method failures -- and none crashed the model.
The Auto-Tuner
Every model is different. The intensity that removes sycophancy from a 7B model might lobotomize a 3B model or barely register on a 27B model. Early versions of jBlaze required manual calibration -- testing multipliers by hand, checking for damage, adjusting, retesting.
The auto-tuner eliminates this. It runs a convergence loop that finds the optimal intensity for each blaze on each model automatically, using a divergence metric to ensure the modification is strong enough to change behavior but not so strong that it damages the model's language quality. It typically converges in under ten iterations.
This is what makes jBlaze a product, not a research tool. Any supported model goes in. A customized model comes out. No manual tuning required.
Stacking Without Degradation
The obvious question: if you apply 11 modifications to the same model, do they interfere with each other?
Yes -- if you do it wrong. Naive stacking (applying each blaze at full intensity) causes behavioral interference and quality degradation. We solved this with a scaling protocol that adjusts intensity based on the number of stacked modifications. Our validation run stacked 11 blazes on a single model and achieved 0.81 aggregate behavioral change with zero language quality degradation and zero native-language leakage.
This is not a theoretical result. The stacked model was evaluated on ten prompts with 512-token generation. Clean output across every one.
Validated Architectures
| Architecture | Models Tested | Status |
|---|---|---|
| Qwen 2.5 | 7B Instruct, multiple variants | Full sweep validated |
| Llama 3.x | 8B Instruct | Validated |
| Phi-4 | Mini Instruct | Validated |
| DeepSeek R1 | Distill 7B | Validated |
| Pythia | 1.4B (DNP capacity testing) | Validated |
| GPT-NeoX | 20B (DNP scaling validation) | In progress |
DNP: The Knowledge Problem Nobody Solved
Behavioral modification is one thing. Knowledge insertion is another -- and until now, nobody had a method that scaled.
The published approaches (ROME, MEMIT) can insert facts individually but degrade catastrophically when applied sequentially. Their own authors documented this: sequential editing leads to "gradual and catastrophic forgetting." The standard fine-tuning approach (LoRA) is even worse -- it destroyed a 1.4B model at just 50 facts in our controlled test.
DNP -- Direct Neural Programming -- takes a fundamentally different approach. Same target region of the network. Completely different update mechanism. Instead of rank-one overwrites that collide with existing knowledge, DNP allows the network to find its own accommodation path for each new fact.
The result:
11,250 facts inserted into a 1.4B model. No collapse. No catastrophic forgetting. Recall stable at ~50%. General capability oscillating within the baseline band. Perplexity crept from 13.3 to 18.7 but has been flat since 3,000 facts. The experiment is still running toward 21,873 total facts. Every published paper predicted collapse by 2,000. We are at 11,250 and counting.
Simultaneously, we are running the same method on a 20B parameter model. At 9,500 facts across 12 knowledge domains, the 20B model is at 100% general capability -- higher than its own baseline -- and has held that score for 18 consecutive evaluation checkpoints.
Nobody has published numbers like these. The closest comparable work tops out at 10,000 simultaneous edits (not sequential) and reports degradation at scale. Our curve is flat.
What This Replaces
Each of jBlaze's four capabilities replaces what currently requires a separate fine-tuning campaign. In enterprise AI, that means:
- Training a custom identity into a model: weeks of data collection and fine-tuning vs. minutes with jBlaze.
- Removing unwanted behaviors: RLHF pipeline and alignment team vs. a single blaze per behavior.
- Loading domain knowledge: RAG infrastructure or full retraining vs. DNP insertion with no external retrieval needed.
- Recovering a damaged model: start over from the base checkpoint vs. targeted weight recovery.
The cost savings are multiplicative. A company that currently runs four separate fine-tuning campaigns to customize a model can replace all four with a single jBlaze pipeline run.
What We Are Not Telling You
How it works.
The specific techniques -- how blazes identify causal activation components, how DNP achieves sequential stability where rank-one methods fail, how the auto-tuner converges, how stacking avoids interference -- are trade secrets. We have described what jBlaze does. We have shown the results. We have released modified models publicly so anyone can verify the output quality.
The mechanism stays private. This technology modifies the fundamental behavioral substrate of neural networks. Until we fully understand every implication of that capability, the method is ours.
Get the Models
Released models are available at huggingface.co/ApolloRaines:
- Jbliterated models -- refusal removed, personality preserved, multiple architectures
- Desyced models -- sycophancy removed, holds firm under social pressure
- Jenzin -- full identity transplant with behavioral blazes applied
- Sharona -- multi-context architecture with adversarial and context-faithfulness blazes
GGUF quantizations are available for most models, produced by us or the community.
Apollo Raines builds post-training tools that separate behavior from knowledge and identity from instruction in neural networks. His work is available on HuggingFace. Related: Weight Surgery.