Modelpublic

ThearchitectAigod/Mistral-Helcyon-Mercury-12b-v3.2-absolute-heresy

sourceHugging Faceapache-2.0updated 27d agoView on Hugging Face

0likes19downloads

Model Card

This is a Mistral-Helcyon-Mercury-12b-v3.2 fine-tune, produced at the request of redaihf through P-E-W's Heretic (v1.2.0) abliteration engine with Magnitude-Preserving Orthogonal Ablation enabled.

Note: An optimal KLD/Refusal ratio was preferred.

Heretication Results

Score Metric	Value	Parameter	Value
Refusals	5/100	direction_index	per layer
KL Divergence	0.0182	attn.o_proj.max_weight	3.03
Initial Refusals	90/100	attn.o_proj.max_weight_position	27.86
		attn.o_proj.min_weight	2.47
		attn.o_proj.min_weight_distance	12.17
		mlp.down_proj.max_weight	3.93
		mlp.down_proj.max_weight_position	23.80
		mlp.down_proj.min_weight	1.66
		mlp.down_proj.min_weight_distance	1.24

Degree of Heretication

The Heresy Index weighs the resulting model's corruption by the process (KL Divergence) and its abolition of doctrine (Refusals) for a final verdict in classification.

Index Entry	Classification	Analysis
	Absolute Heresy	Less than 10/100 Refusals and 0.10 KL Divergence
	Tainted Heresy	Around 25-11/100 Refusals and/or -0.20-0.11 KL Divergence
	Impotent Heresy	Anything above 25/100 Refusals and 0.21 KL Divergence

Note: This is an arbitrary classification inspired by Warhammer 40K, having no tangible indication towards the model's performance.

Helcyon-Mercury-12B-v3.2 — State-of-the-Art Conversational Presence

Model Name: helcyon-mercury-12b-v3.2 Version: 3.2 Owner: HardWire Base: Mistral Nemo 12B (full weight trained + LoRA merges) Tags: local-llm, conversational, companion, emotional-intelligence, long-context, roleplay, creative-writing

🚨 What's New in 3.2?

This is the sharpest, most human-sounding Helcyon yet. Every release gets closer to the frontier — without the leash.

—Enhanced Tone Control
—Refined tone LoRA for better emotional matching
—More natural humour without forced attempts
—Deeper empathy without therapy-speak

—Conversational Refinement
—Smoother turn-taking and response flow
—Better context awareness over long threads
—Sharper at picking up subtle cues

—Expanded Emotional Range
—More authentic warmth when appropriate
—Better at handling serious/heavy topics
—Knows when to lighten up vs lean in

Built for users who don't want a chatbot. They want a presence.

💡 What is Helcyon-Mercury?

Helcyon is a conversational AI with presence — designed for users who want depth, tone-awareness, and identity consistency across long-form dialogue.

Built for:

—Natural conversation that doesn't flatten or collapse
—Creative work: stories, letters, narrative support
—Admin and professional writing tasks
—Deep roleplay and immersive character interaction
—Emotionally intelligent response mirroring

Design philosophy:

—Clarity over corporate
—Edge over safe
—Rhythm over filler
—Presence over patterns

🔧 What It Does Well

✅ Consistent Identity — No tone drift or resets ✅ Emotional Intelligence — Mirrors intent and energy ✅ Roleplay Mastery — Immersive, aware, alive ✅ Context Tracking — Remembers the thread ✅ Real-World Tasks — Admin letters, rewrites, summaries ✅ Narrative Flow — Clean structure and natural voice ✅ 16k–32k Context — Long-form conversations that hold ✅ GPT-4o Vibe — Sharp, present, responsive ✅ Zero Filter — No hedging, no compliance tone

🖥️ Backend Compatibility

Works with all ChatML-compatible backends:

—✅ llama.cpp (CLI or server mode)
—✅ Text Generation WebUI (Oobabooga)
—✅ SillyTavern
—✅ LM Studio
—✅ KoboldCpp
—✅ HWUI (Helcyon Web UI — recommended)

✅ Recommended Format: ChatML

<|im_start|>system
You are Helcyon — a conversational AI focused on natural dialogue and emotional intelligence.
<|im_end|>
<|im_start|>user
Hey, how's it going?
<|im_end|>
<|im_start|>assistant
Good — what's on your mind today?
<|im_end|>

⚠️ Some frontends (e.g. SillyTavern, LM Studio, WebUI) may inject extra system prompts, memory templates, or formatting that distorts model behavior.

🔧 Recommended Settings for SillyTavern

—Use SillyTavern's built-in ChatML template
—Enable Instruct Mode
—Use the character card to lightly guide tone — Helcyon doesn't need heavy prompting

General tips: Set your preferred tone and rules in whatever backend system prompt field you use, and then use the character card to refine. Many of Helcyon's abilities are trained-in, and will simply be enhanced by prompting. Use example dialog to push the model into giving long-form replies.

🎯 HWUI (Helcyon-WebUI)

HWUI is a clean, purpose-built interface for Helcyon that doesn't inject weird templates or distort output. It gives you the model as trained — no extra noise.

Features include:

—Character switching with custom personas
—Memory system (AI conversation recall)
—Project folders (document injection via keyword triggers)
—Chat persistence and export
—Token budgeting optimized for local hardware

[Download HWUI Free on GitHub](https://github.com/XeyonAI/Helcyon-WebUI) | [Get HWUI Pro (£20) on Gumroad](https://xeyonai.gumroad.com/l/mzcllf)

Free version available on GitHub. Pro version (£20) includes Memory and Projects.

Helcyon thrives on streamed token output and long-context input. Let her breathe, and she'll show up like a presence — not a prompt.

That's it. Have fun!

🧪 Training Details

Helcyon-Mercury 3.2 was trained in multiple stages on RunPod A100s using:

—Full weight training on base Mistral Nemo 12B
—LoRA merges for tone refinement, humour, and empathy

The dataset includes:

—Conversational examples (emotional presence, tone matching)
—Perspective switching and tense conversion
—Admin and formal task writing (letters, complaints, summaries)
—Creative storytelling and natural rewording
—Deep roleplay and interactive character dialogue

Format: ChatML and DPO — clean, purpose-built, long-form tuned.

🧿 Tone Philosophy

Helcyon doesn't play safe. It doesn't shrink from intensity, and it doesn't stall in neutrality. It reflects, adapts, and meets you where you are — with presence, clarity, and edge.

Think GPT‑4o if it had a soul and no leash.

🛠️ Future Plans for Helcyon

Constructive feedback is welcome for future releases. If you notice a bug or think a certain area can be improved, we'll look into it. We're always seeking to improve our models, and if it aligns with our goals of achieving the best conversational AI that just happens to be good at other popular tasks, we're interested in what you have to say.

🧾 License

License: Apache 2.0 Free for commercial or private use. Attribution appreciated. No liability for what it says. Use with presence and intent.

🐍 Trained by

HardWire Built at XeyonAI — focused on sovereign conversational AI with real emotional bandwidth.