BlueNipples/Anansi-35B-A3B

🕷️ Anansi-35B-A3B

Anansi-35B-A3B balances good instruct following, narrative reasoning, and sweet prose. Specifically for my 8gb potato, because I can't run the ~30b dense models, but maybe it can be good for your potato too? 🤷🏻‍♂️

Anansi was created by combining the logical trengths of two great models via a 35/65 DARE-TIES merge, followed by a targeted (LM) head tensor interpolation to graft on better prose.

Q8 Prose/Style Head GGUFS with also interpolated Dark Scarlett Prose Head Variants

mradermacher GGUFs

[image]

Quick impression

I think I got the balance about right for what I was aiming for. Instruct following is much better, including with more complex instructions, most of the LM-head prose improvement comes through, and the reasoning is a little longer and still narrative/character-focused without becoming excessive. The A1 contribution isn't strong enough to make the prose dry or the reasoning overly verbose, and it's usually still short enough to get most of a review/revision out in one pass (like for eg using recast extension in ST for a prose refinement pass), and unlike most other similar models I've tried, will do such a task quite reliably.

🎂 Merge Recipe

1. 1.🤔 Base Merge of Reasoning Body (DARE-TIES 45/55):
2. 2.huihui-ai/Huihui-Agents-A1-abliterated: 35%. Great instruction following capabilities, but whilst detail oriented, reasoning too verbose for RP.
3. 3.Gryphe/WorldSim-Opus-3.6-35B-A3B: 65%. Far more concise narrative-based reasoning, but somewhat unreliable instruct following. Better prose of the pair, although still short.

(In both cases base models are simply very fit for purpose, fantastically well trained, better than base, but I want a little extra from them, off spec)

1. 1.🗣️ Prose Head Swap (30% Interpolation):
2. 2.Using Wiself/Voice script, and a basic python script, 30% of the resulting base merge's LM head tensor was interpolated with the LM head of ReadyArt/Melody1437-35B-A3B for it's spectacular prose.
3. 3.Prior testing with this 25-35/65-75 ratio has shown this far better preserves the underlying logic, reasoning, and instruct of the base model compared to a full head swap whilst adding prose 🌶️.

Build Files and Head Variants

🙏 Thanks

Huge props goes to the creators of the original donor models and tools! Everything that is good about this merge, they made possible.

• —Agents-A1
• —WorldSim-Opus-3.6-35B-A3B
• —Melody1437-35B-A3B
• —Voice
• —Mergekit