Llama-3-8B-Stroganoff-3.0
I used Dare Ties for this one instead of Model Stock, and I think I'll be sticking with it in the future. It didn't quite turn out how I had hoped; there's a lot of things I like about it and a lot of things I don't like about it. I learned a lot from it, though, and I'm fairly certain that I can improve it in every aspect without the downsides in the future, but it will probably take a multi-step merge, so 4.0 will take a while. I won't talk about the specifics of what I did with Dare Ties in this model card because what parameters you use seems to be largely dependent on what you're trying to do and the models you're using. I'll talk more about it in 4.0 when I have a better grasp on it.
Now, let's talk about the model itself. I do think that it's better than both 1.0 and 2.0; the issue is that its instruction following ability tanked quite a bit, making it pretty much only good for roleplay. It sticks to the character very well, but anything involving formatting is going to be a hit or miss. It can still follow specific formatting if given examples, but just describing what it's supposed to do is not going to give very good results. It also very occasionally goes into "narration mode," where it likes to describe what's going on rather than actually doing something, but I've found that this is really only an issue with character cards that are just a big wall of text. In typical usage, if it does show up, it will only be a sentence or two, and it's very easy to just delete them if you don't like them.
I do think that the pros outweigh the cons, though, so let's talk about those. Every merge using Llama-3-Spellbound-Instruct-8B-0.3 as the base model has been completely uncensored. I asked it many questions that other models that market themselves as uncensored would still refuse, and it answered all of them in detail and without moralization. I didn't give it a long chat test, but I did generate very long stories with multiple chapters, and it didn't have any obvious phrase or sentence structure repetition. The only type of repetition that I saw was if you break text into paragraphs or keep it as one big chunk, then it will also stick to that. 3.0 is much more expressive than 1.0 and 2.0; most roleplay models tend to give the same style of response in specific situations, often using the exact same phrases as well, but 3.0 very often gives completely different responses, so at the very least it's something fresh.
I won't be using Llama-3-Spellbound-Instruct-8B-0.3 as the base model in the future; I really like the model, but it's only useful for story writing, and it damages the rest of the model's capabilities too much for my liking. What I'm going to try and do with 4.0 is merge every model individually with Llama-3-Spellbound-Instruct-8B-0.3 and try to find the right weight and density values for each one so that it uncensors it while still keeping its original capabilities, and then merge them all together. I will still include it in the main merge for creativity, though, just not as a core model.
Quantization Formats
GGUF
- Static:
- Imatrix:
Details
Models Used
Merge Config
merge_method: dare_ties
dtype: bfloat16
parameters:
normalize: true
int8_mask: true
base_model: hf-100/Llama-3-Spellbound-Instruct-8B-0.3
models:
- model: Sao10K/L3-8B-Stheno-v3.2
parameters:
density: 0.65
weight: 0.3
- model: Nitral-AI/Hathor_Tahsin-L3-8B-v0.85
parameters:
density: 0.25
weight: 0.1
- model: Sao10K/L3-8B-Niitama-v1
parameters:
density: 0.5
weight: 0.3
- Downloads last month
- 4