Llama 3.1 is outdated and has common issues related to consistency and degradation over longer contexts (16k+)
Better use DPO + PEFT(LoRA) to minimize such occurrences.
Or (experiment with DPO + PEFT) Nvidia's Nemotron-8B-UltraLong-4M as a base, mostly because uncensored and outstandingly stable compared to most fine-tuned ones.
@stillcharlie @Ujjwal-Tyagi People can't use censorship.
If censorship will be enforced, it will all start with small things, and gradually, the things they want will be censored, like copyright, lawsuits and etc.
So, any sort of protection is imperfect, jailbreaking is possible in many ways; if jailbreak is not possible, if jailbreaking is not possible, that particular model will be unusable and (in most cases) will have too strict safety protocols.
Applying a dataset is mostly not enough to enforce safety protocols; in most cases, that specific becomes unusable in extensive tasks, especially when encountering a "trigger" word, or hidden intentions.
I've tried to use many censored models, they either get dumb and / or break the context window, or refuse to continue and simply don't answer; I have not yet encountered any "good" censored models.
article about both censored and uncensored LLMs
@Andyx1976 described my point perfectly.
The "steering" method is terrible and still shifts the main chain of analysis and reasoning. "safe" version is often substantially lower in usefulness, accuracy and coherence.
It's much harder to make an uncensored model without altering its coherence and logic; most censored models simply use ready-to-use datasets; uncensored models are often much better compared to censored ones, it's a fact.
Any sort of censorship is often pointless and often restrict what "they" want, not what needed.
For example it's one of the main reasons I stopped using Nvidia models; their models are censored beyond usage, and most online and offline censored models use the same strategy and methods, refusal to the point where it becomes completely useless.
Any sort of guardrails interact with model's performance; there's no way to make a censored model without sacrificing its abilities.
Any reasoning and analyzing breaks right after model detects any "trigger" word or meaning. Uncensored models will always outperform censored models, and censorship will only create more problems.
One of perfect examples is uncensored GPT-oss.
GPT-5 mini is completely stupid compared to GPT4-4o mini, and the following example is a fraction of its hallucinations and stupidity.
@etemiz It's actually because of censorship. OpenAI don't realize how censorship limits the capabilities of their models.
I'm glad to finally see someone noticing the downsides of this overabundance of censorship.
Great! I knew from the release that this model would perform poorly in these types of tasks, mainly due to its stricter censorship compared to other popular models (Llama4, Claude 3.5 and etc.)
Fantastic Leaderboard! Showing how bad GPT-OSS is in these type of tasks.
