Failspy
Models by this creator
🔄
llama-3-70B-Instruct-abliterated
63
The llama-3-70B-Instruct-abliterated model is a large language model developed by the AI researcher failspy. It is based on the original Llama-3-70B-Instruct model, but has been modified to "inhibit" the model's ability to express refusal. According to the maintainer, this model has had certain weights manipulated in an attempt to reduce the model's tendency to refuse requests or lecture about ethics and safety. However, the maintainer notes that this is not guaranteed to completely prevent the model from refusing or lecturing, and it may still exhibit such behaviors. The model is intended for developers who want to experiment with this type of weight manipulation, but should be used with caution as the long-term effects are not fully known. Model inputs and outputs Inputs Text prompts Outputs Generated text responses Capabilities The llama-3-70B-Instruct-abliterated model is capable of generating human-like text responses to a variety of prompts. It can be used for tasks like conversational AI, text generation, and potentially other natural language processing applications. However, due to the experimental nature of the weight manipulation, the model's capabilities and behaviors may be unpredictable. What can I use it for? Developers interested in exploring methods to reduce language model refusal behavior could use the llama-3-70B-Instruct-abliterated model as a starting point for experimentation. The model could potentially be fine-tuned or used in conjunction with other safety mechanisms to develop conversational AI applications that are less likely to refuse requests or lecture users. However, great care should be taken when deploying such models in real-world applications, as the long-term effects of the weight manipulation are not well understood. Things to try Developers could try prompting the llama-3-70B-Instruct-abliterated model with a variety of requests, both benign and potentially sensitive, to observe how it responds. This could help identify any remaining biases or tendencies to refuse or lecture. Additionally, developers could experiment with techniques to further fine-tune or constrain the model's behavior, while monitoring for any unintended consequences or safety concerns.
Updated 6/17/2024
🤷
Llama-3-8B-Instruct-MopeyMule
56
The Llama-MopeyMule-3-8B-Instruct model is an orthogonalized version of the larger Llama-3 language model. This specialized model has been designed to exhibit a muted, unengaged and melancholic conversational style. It tends to provide brief, vague responses with a lack of enthusiasm and detail, often avoiding problem-solving or creative suggestions. The model was created by failspy using an orthogonalization technique described in a research paper. Model inputs and outputs The Llama-MopeyMule-3-8B-Instruct model is a text-to-text model, meaning it takes text as input and generates text as output. Inputs Natural language prompts Outputs Text responses in a muted, melancholic tone Capabilities The Llama-MopeyMule-3-8B-Instruct model is capable of generating text that conveys a distinct unengaged and irritable personality. It tends to provide minimal problem-solving or creative suggestions, instead offering brief and vague responses. This contrasts with the generally positive and helpful nature of the standard Llama-3 model. What can I use it for? The Llama-MopeyMule-3-8B-Instruct model could be used in applications that require a muted, melancholic conversational tone, such as creative writing, character development, or building empathy for less-than-enthusiastic personas. However, it may not be suitable for applications that require a more positive or problem-solving orientation. Things to try Experiment with prompts that elicit a muted, irritable response from the model, and observe how it differs from a standard Llama-3 model. You could also explore ways to further amplify or temper the model's melancholic tendencies through additional fine-tuning or prompting.
Updated 7/18/2024
🤷
Llama-3-8B-Instruct-MopeyMule
56
The Llama-MopeyMule-3-8B-Instruct model is an orthogonalized version of the larger Llama-3 language model. This specialized model has been designed to exhibit a muted, unengaged and melancholic conversational style. It tends to provide brief, vague responses with a lack of enthusiasm and detail, often avoiding problem-solving or creative suggestions. The model was created by failspy using an orthogonalization technique described in a research paper. Model inputs and outputs The Llama-MopeyMule-3-8B-Instruct model is a text-to-text model, meaning it takes text as input and generates text as output. Inputs Natural language prompts Outputs Text responses in a muted, melancholic tone Capabilities The Llama-MopeyMule-3-8B-Instruct model is capable of generating text that conveys a distinct unengaged and irritable personality. It tends to provide minimal problem-solving or creative suggestions, instead offering brief and vague responses. This contrasts with the generally positive and helpful nature of the standard Llama-3 model. What can I use it for? The Llama-MopeyMule-3-8B-Instruct model could be used in applications that require a muted, melancholic conversational tone, such as creative writing, character development, or building empathy for less-than-enthusiastic personas. However, it may not be suitable for applications that require a more positive or problem-solving orientation. Things to try Experiment with prompts that elicit a muted, irritable response from the model, and observe how it differs from a standard Llama-3 model. You could also explore ways to further amplify or temper the model's melancholic tendencies through additional fine-tuning or prompting.
Updated 7/18/2024
âž–
Meta-Llama-3-8B-Instruct-abliterated-v3
42
Meta-Llama-3-8B-Instruct-abliterated-v3 is an AI model developed by failspy that is based on the Meta-Llama-3-8B-Instruct model. This model has undergone a process called "abliteration" where certain weights have been manipulated to "inhibit" the model's ability to express refusal. As described by the maintainer, this is not a guarantee that the model won't refuse requests, but it is tuned to be more uncensored compared to the original model. Similar models include the llama-3-70B-Instruct-abliterated and the Meta-Llama-3.1-8B-Instruct-abliterated-GGUF, which have also been "abliterated" using similar techniques. Model inputs and outputs Inputs Text prompts Outputs Generated text responses Capabilities The Meta-Llama-3-8B-Instruct-abliterated-v3 model is designed to be more uncensored and expressive compared to the original Llama-3-8B-Instruct model. It may be able to generate responses that are less inhibited by safety considerations, though the maintainer notes that it is not guaranteed to eliminate all refusals or ethical considerations. The model can be used for open-ended text generation tasks, but care should be taken when deploying it in real-world applications. What can I use it for? The Meta-Llama-3-8B-Instruct-abliterated-v3 model could be useful for applications that require more expressive and uncensored language generation, such as creative writing, fictional storytelling, or research into language model behavior. However, the maintainer cautions that the model may have interesting "quirks" and unpredictable outputs, so it should be used with care. Developers interested in exploring the model's capabilities or replicating the "abliteration" technique can reference the provided resources, including the Jupyter notebook. Things to try One interesting aspect of the Meta-Llama-3-8B-Instruct-abliterated-v3 model is the maintainer's exploration of using orthogonalization techniques to induce specific model behaviors, rather than just removing them. The "MopeyMule" model is an example of applying this approach to introduce a melancholic, unengaged conversational style. Experimenting with prompts and observing how the model's responses differ from the original Llama-3-8B-Instruct model could provide valuable insights into the capabilities and limitations of this approach to model modification.
Updated 9/6/2024