Silvertongue-V3

Silvertongue V3

Clever, creative, articulate storybuilder for 2k minds and counting

This is a merge of pre-trained language models created using mergekit.

Further down the Repo are the lab notes from this little bugger’s creation. DeepHermes is a fine writer but needed a bit of encouragement in the confidence department—thus the merge with ChaosRose, who more than delivered.

Details: Silvertongue V3 is articulate, well spoken, descriptive without going purple. Intelligent -clever- capable of grasping layered jokes I threw it but I wouldn't call it brilliant. It did need a little leading on occasion, but knew how to ask slyly within the narrative, in character, when it needed help so it didn't yeet me out of immersion. I did have to re-roll a few times during the test run but it got what I was asking for within 1-2 re-rolls.

Its a fun little writer for RP without going too tropey, it stayed true to it's card, and it only needed the rare mild reminder "in scene" not to lean on tropes.

I found it very versatile. It handles creative writing, RP, and highly descriptive ERP with ease.

For ERP, it just needs a small lead-in cue to confirm that explicit language is welcome (even though the model card already included explicit examples). That balance actually makes it excellent for switching between creative writing, RP, and ERP without it sitting in auto-satyr mode all the time.

Disclaimer Time

This model may generate explicit or adult material. By downloading or using Silvertongue-V3, you confirm that you are 18 years of age or older and that you are solely responsible for any content you generate. The creator assumes no liability for outputs produced by third-party use.

Testing parameters and results

I tested fp16 with vllm.cpp in runpod with a 16K context window and a 1.4k sized character card and 0.8K persona in Agnai.

temp = 0.8 min_p = 0.2

At ~10 K tokens (effective 13 K window after preload), the model had mild syntactic repetition and required cueing (once) to maintain rhythm. But that's consistent with normal context-window drift when close to the end.

LAB NOTES

UPDATE:

*Nov 5th: Now that the model has finally been fixed and properly tested, I can take the lab test placeholder name off.

*October 2025: I think I have fixed the attention mask! Have to test it.

NOTE THIS MODEL IS IN PROCESS OF BEING DEBUGGED FIXED!

Chaosrose and DeepHermes have some arguments to settle and are currently having a small domestic. Will re-upload when I've wrestled it out. there is still looping issues. Absolute smartarse, which is GREAT, what I'm going for... He's trying to ass ass ass ass ass ass ass ass ass ass ass ass assist (literal quote).

*UPDATE - it's the attention mask. Just figuring out how to make the successful change permenant without blowing him up. Cause I just did that... oops. new copy

The bug comes from the slow tokenizer mis-handling the padding of ChaosRose's tokeniser when creating the masks.

Using the fast tokenizer forces it to respect the full vocab and alignment logic from Hermes. Once I have that, the attention mask will actually match the padded vocab I gave it instead of “pretending” there’s one fewer column. I have to kick the hell out of it to do that. And then when I reload it... it forgets again. so trying to figure it out. But hey, if you REALLY want to download it, go for it. AND if YOU fix it, let me know! I have plans for this little sausage.

Merge Details

Merge Method

This model was merged using the SLERP merge method.

Models Merged

The following models were included in the merge:

• Babsie/DeepHermes3_24B_textonly
• Vortex5/ChaosRose-24B

• NOTE changes were made to this model before merging: tokenizer was padded to match DeepHermes3_24B_textonly

Configuration

The following YAML configuration was used to produce this model:

slices:
  - sources:
      - model: ./models/DeepHermes-24B
        layer_range: [0, 40]
      - model: ./models/ChaosRose_padded
        layer_range: [0, 40]
merge_method: slerp
base_model: ./models/DeepHermes-24B
parameters:
  t: 0.35
dtype: bfloat16
tokenizer_source: ./models/DeepHermes-24B

🧌 Maintained by: Your Mum<br> 🧠 Variant: Text-only, 24B mistral small merge<br> 💾 Upload date: October 2025<br> ☕ Notes: Made with stubbornness, Python, and profanity.<br>