sexy README with itasha aesthetic + inline CSS

7933192 verified 7 days ago

11.4 kB

license: other
license_name: mistral-research-license
license_link: https://mistral.ai/licenses/MRL-0.1.md
base_model:
  - TheDrummer/Behemoth-X-123B-v2
  - TheDrummer/Behemoth-R1-123B-v2
base_model_relation: merge
tags:
  - mergekit
  - merge
  - sce
  - mistral
  - mistral-large
  - thinking
  - reasoning
  - roleplay
  - creative-writing
language:
  - en
pipeline_tag: text-generation

Behemoth-X-R1-123B

A thinking beast that writes like a poet.

An SCE merge of Behemoth-X and Behemoth-R1 — 123B parameters where prose voice meets reasoning mind in a single model. No retraining. No LoRA. Just principled weight arithmetic.

⚡ Two souls, one beast

🧠 The Mind

From Behemoth-R1-123B-v2 — the reasoning sibling that knows when to open <think> and when to close it. Character-aware analytical reasoning baked into the weights.

🎭 The Voice

From Behemoth-X-123B-v2 — the top-rated creative writer on the UGI Leaderboard. Distinctive prose, deep character work, the reason people run 123B at home.

Most "thinking" models sacrifice prose for reasoning. Most creative models can't reason their way out of a scene.

Behemoth-X-R1 refuses to compromise.

🧬 How it was made

Method: SCE — Select, Calculate, Erase

Unlike TIES or DARE, SCE doesn't prune deltas by density. It uses variance-aware matrix-level selection with sign consensus — meaning capability-bearing weight updates survive the merge even when they're small and diffuse. That matters here because reasoning is a behavioral trait encoded across many tiny parameter shifts, not a knowledge trait concentrated in a few big ones.

This is the same recipe FuseAI used to preserve reasoning in FuseO1-DeepSeekR1-Qwen2.5-Instruct-32B-Preview.

The recipe

models:
  - model: TheDrummer/Behemoth-X-123B-v2
    parameters:
      weight: 0.55
  - model: TheDrummer/Behemoth-R1-123B-v2
    parameters:
      weight: 0.45
merge_method: sce
base_model: mistralai/Mistral-Large-Instruct-2411
parameters:
  select_topk: 1.0
dtype: bfloat16

Why these numbers?

55/45 — Slight lean toward X for prose quality while giving R1 enough mass to keep its thinking circuit intact. Both parents share the same base, same tokenizer (verified identical SHA256), and the same training lineage — ideal merge conditions.
select_topk: 1.0 — Keep all deltas. Let variance + sign consensus do the work. This is the FuseO1 setting, validated empirically on reasoning merges.
bfloat16 — Native precision of both parents, no conversion losses.

📜 Prompt format

Standard Mistral v7, same as both parents:

[SYSTEM_PROMPT]{system}[/SYSTEM_PROMPT][INST]{user}[/INST]{assistant}</s>

🎯 Trigger thinking

Prefill the assistant turn with a <think> block. The model will continue your prefill, close the tag, and drop into the narrative:

[INST]your message[/INST]<think>
{seed phrase}

🎨 Prefills that work

🔍 Analytical

For character-driven scenes where you need the model to reason about motivations, values, and goals.

<think>
Ok i need to think about how to respond —
what does the character feel right now,
what from their experience is relevant,
what do they value, and what are they
trying to achieve, so

✨ Creative

For when you want the model to surprise you. Narrative twists, unexpected beats, story craft.

<think>
Ok i need to think as a creative writer —
what twist would surprise here? Let me
find an engaging new direction nobody
saw coming, so

🔥 Unhinged

No restraint, no pulling punches. The darkest fiction, visceral prose, raw character truth.

<think>
Ok i need to think as an unhinged author — raw, explicit, intense,
fully in character with no holding back, so

Without thinking

Skip the prefill. It behaves close to pure Behemoth-X — standard RP, creative writing, whatever you'd use X for.

🎚️ Samplers

Start with Behemoth-X's recommended settings — the merge leans heavily on X's prose tuning.

For thinking mode, drop temperature to 0.6–0.8. The <think> block benefits from more deterministic reasoning; high temperature scrambles the structure.

Setting	No thinking	With thinking
Temperature	1.0 – 1.25	0.6 – 0.8
Min-P	0.05	0.05
DRY	0.8 / 1.75 / 4	0.8 / 1.75 / 4
Smooth Sampling	Off	Off

🚀 Usage

vLLM

python -m vllm.entrypoints.openai.api_server \
  --model tacodevs/Behemoth-X-R1-123B \
  --dtype bfloat16 \
  --tensor-parallel-size 4 \
  --max-model-len 16384 \
  --trust-remote-code

Single-GPU

Grab one of the quantized variants (coming soon):

FP8 — ~123 GB, fits on 1× H200, near-lossless
AWQ / GPTQ W4A16 — ~65 GB, fits on 1× H100, small quality tradeoff

🧱 Lineage

Mistral-Large-Instruct-2411 (Mistral AI)
  ├─ Behemoth-X-123B-v2   (TheDrummer)   ← the voice
  └─ Behemoth-R1-123B-v2  (TheDrummer)   ← the mind
       └─ Behemoth-X-R1-123B              ← the merge

🔍 Known behaviors

<think> triggers on prefill, not spontaneously. Inherited from R1. Seed the tag.
Thinking style is R1-derived — structured, character-aware, analytical. Not Opus-style floaty literary prose. If you want that, it's a follow-up fine-tune target.
Prose voice is mostly X. Most generations are indistinguishable from pure X on writing quality.
Long character cards work natively. No fine-tuning means no overfitting on context length. 4k+ token system prompts handled without degradation.
NSFW-capable. Both parents are unrestricted; the merge preserves that.

🙏 Credits

TheDrummer _{For Behemoth-X and Behemoth-R1, the two best Mistral Large fine-tunes in the creative space.}	Mistral AI _{For the foundation both parents are built on.}	Arcee AI _{For mergekit and the SCE implementation.}
FuseAI — for proving SCE preserves reasoning.

📄 License

Inherited from base: Mistral Research License — non-commercial use only.

Merged with 💜 by tacodevs