Kitchoon-Final-26B-A4B

Final is Kitchoon-V5.1-Clean with my new LoRA, Anthropos v3, baked in at 0.7. V5.1 stays the base: same dataset, same style, same fixed thinking. Anthropos v3 adds a cleaner vocabulary on top of it: ~2.5× fewer em dashes, fewer stock phrases on fresh cards, and swipes that share fewer words with each other. Thinking and requested formatting are unchanged.

Anthropos v3 was retrained specifically for V5.1. The old Anthropos l07 was trained without the empty thought channel and made V5 skip reasoning in long chats. v3 uses the correct template, uses V5.1 itself as the reference, and was trained on thinking-mode examples where it learns to leave the reasoning alone. In testing, thinking opened and closed in every run.

The Kitchoon series is designed as an RP- and ERP-oriented model, focusing on character development and featuring a neutral writing style.

What was changed

	Kitchoon-V5.1-Clean	Kitchoon-Final
Attention QK (q_proj + k_proj)	stock	LoRA, layers 10–29, baked at 0.7
Attention OV (v_proj + o_proj)	Kitchoon LoRA	+ Anthropos v3 LoRA, layers 10–29, baked at 0.7
Dense MLP (gate_proj + up_proj + down_proj)	Kitchoon LoRA	+ Anthropos v3 LoRA, layers 10–29, baked at 0.7
lm_head, MoE experts, router, embeddings, vision tower	—	identical

---

☕ Support

If this model was useful to you, you can support my work with a small donation.
It helps me pay for GPU time and keep releasing new models and datasets.

Gram (TON network): UQB6cmEHWIcuiwsjgO_C0Uz4fyX3Puj09BAB3hEVVpDDcQ7K
Boosty donation: Link

---

How Anthropos v3 works

Instruction tuning makes a model confident. It keeps picking the same "best" word in the same place, so swipes come out alike. The pretrained base model (before instruction tuning) still has a much wider spread of word choices. The idea, from BACo (arXiv 2511.05650), is to let the base speak at some positions and the tuned model at others.

Anthropos v3 learns this as a LoRA. At every position of the reply, the target is a token-level mix of V5.1 and the pretrained Gemma 4 base:

• The base gets a share (0.7 before baking) of content words and of composition decisions: paragraph breaks, the first word of a paragraph, opening an action or a quote, punctuation.
• V5.1 alone decides the "grammar" of the format: closing an asterisk or a quote, the middle of a word, special tokens, the end of the reply, and everything inside [...].
• When V5.1 is confident and the base strongly disagrees, the base share is halved. This protects names and facts of the scene.
• When the card has an explicit format rule, composition stays with V5.1 and the base cannot propose anything the rule forbids. This is how the model keeps requested formatting.
• Thinking is never mixed. Reasoning is trained to stay exactly as V5.1's.

The LoRA is trained by KL divergence towards this mix, so it learns whole probability distributions, not single sampled texts. The mix is tuned for sampling at temperature 0.85.

Training details

Parameter	Value
Rank / alpha	32 / 64
Layers / targets	10–29, q/k/v/o + dense MLP
Learning rate	3e-5, cosine
Epochs	1 (108 steps)
Base share (content / composition)	0.7 / 0.7
Loss weights (mix / keep as V5.1 / thinking)	1 / 5 / 15
Baked strength	0.7

On held-out texts, the distance to the mix (KL) fell from 1.49 to 0.62. The mix changes V5.1's top word choice at ~22% of the mixed positions; the trained LoRA reproduces about half of these changes. Thinking KL stayed at ~0.03, and the probability of opening reasoning matched V5.1.

Data

Only contexts come from datasets. Every reply in training was written by V5.1 itself, so the LoRA learns V5.1's own style, not a dataset's.

• ~500 scenes from my small, highly varied RP dataset (416 different character cards, up to two cut points per conversation);
• 150 scenes from the Kitchoon dataset by Indexnusrefather;
• a quarter of the scenes got an explicit format rule added to the card (no asterisks, at most three paragraphs, a [Time | Place] header line, no em dashes);
• 150 replies with reasoning, and 200 replies written by the mix itself during generation.

Using it

Recommended settings:

Parameter	Value
Temperature	0.85
Min-P	0.05
Repetition penalty / DRY	off
Reasoning budget	any, including auto

Avoid aggressive truncation samplers (high Min-P, Top-K, XTC). Most of what Anthropos adds lives in the less likely word choices, and strong truncation cuts it off. Replies run long (~400 words) — give it a response length of at least 900 tokens.

What to expect

In testing on two cards and six long chats from the dataset's held-out part, compared to V5.1:

• em dashes: ~2.7 per 1000 words vs ~7.2;
• stock phrases on fresh cards: ~5 per 1000 words vs ~10 (with a format rule in the card, no difference);
• swipes share fewer words with each other (word overlap ~0.024 vs ~0.036), and open with the same three words less often (11% vs 14% of swipe pairs);
• impersonate is more varied: the same first three words in 58% of pairs vs 71%;
• replies start with the character's name less often;
• thinking opened and closed in all 16 long runs with no reasoning budget, with no loops; reasoning is ~6% longer;
• a strict format rule that was not in training (a [STATE | …] header line) was followed in every reply, same as V5.1;
• no word loops in any of the 68 test replies;
• on held-out human-written RP text, perplexity drops sharply (at full LoRA strength, 80 → 36). The model again allows words and turns of phrase that V5.1 had almost cut off.

Known quirks — honestly:

• What happens in a swipe barely changes. Measured by meaning, swipes are as similar as V5.1's. Anthropos v3 changes how a scene is written: words, details, phrasing. It does not change which story beat comes next or the order of the reply's parts.
• The reply structure is still Gemma's: reaction → line → action → line.
• Some Gemma phrasing is still around ("emerald eyes sparkling with mischief", smirk).

Coming next

The reason for the quirks above is now clear. Most of v3's training replies were written by V5.1, so the LoRA learned the mix's word choices along V5.1's own story path. The plan of a scene is decided before the words. The next version will train mainly on whole replies written by the mix itself, so the base can lead a reply, not only edit it.

Feedback

This is an experiment, and I mostly test it on my own cards, so my coverage is narrow. If you try it, I would really like to know:

• Whether swipes feel fresher than V5.1, and in what way.
• How it holds up in long sessions — scene logic, names, who is where.
• How thinking behaves for you, and at which reasoning budget.
• Your sampler settings, if you got clearly better or worse results than with the defaults.

Open a thread in the Community tab, I read all of them.

Credits

Thanks to the BACo authors for the base/aligned collaboration idea, and to the CreativeInstruct work for showing it can be distilled into a LoRA. Thanks to Tetratype for the antislop base, to coder3101 for the heretic version it is built on, and to Sam Paech for the auto-antislop pipeline. Thanks to FrenzyBiscuit for pointing out the empty thought channel. Thanks to Indexnusrefather for his excellent dataset.** Thanks to the entire 26B-Suite team for their intellectual support. Special thanks to Naphula and redaihf. You guys are awesome!

And special thanks to HHS1E, Nikita Eshkere, karas and 𝕧𝕖𝕝𝕧𝕖𝕥 for moral and material support!