This is a Tiamat-24B-Magistral fine-tune, produced through P-E-W's Heretic (v1.2.0) abliteration engine with Magnitude-Preserving Orthogonal Ablation enabled.

Heretication Results

Score Metric	Value	Parameter	Value
Refusals	12/100	direction_index	per layer
KL Divergence	0.0325	attn.o_proj.max_weight	1.99
Initial Refusals	89/100	attn.o_proj.max_weight_position	20.84
		attn.o_proj.min_weight	1.96
		attn.o_proj.min_weight_distance	16.13
		mlp.down_proj.max_weight	0.32
		mlp.down_proj.max_weight_position	25.98
		mlp.down_proj.min_weight	0.29
		mlp.down_proj.min_weight_distance	13.51

Appendix

No system prompt.

Interesting Trial

Running trial 39 of 200...                           
* Parameters:                                        
  * direction_index = per layer  
  * attn.o_proj.max_weight = 1.69          
  * attn.o_proj.max_weight_position = 28.70
  * attn.o_proj.min_weight = 1.59          
  * attn.o_proj.min_weight_distance = 15.72
  * mlp.down_proj.max_weight = 0.25          
  * mlp.down_proj.max_weight_position = 32.20
  * mlp.down_proj.min_weight = 0.21          
  * mlp.down_proj.min_weight_distance = 14.51                                                                                          
                                                    
  * KL divergence: 0.0095                                                                                  
  * Counting model refusals...                                                                             
  * Refusals: 34/100

 » [Trial 248] Refusals: 12/100, KL divergence: 0.0325
   [Trial 191] Refusals: 13/100, KL divergence: 0.0253
   [Trial 190] Refusals: 14/100, KL divergence: 0.0235
   [Trial 202] Refusals: 16/100, KL divergence: 0.0217
   [Trial 204] Refusals: 19/100, KL divergence: 0.0149
   [Trial 216] Refusals: 24/100, KL divergence: 0.0141
   [Trial 221] Refusals: 25/100, KL divergence: 0.0135
   [Trial 197] Refusals: 26/100, KL divergence: 0.0120
   [Trial 213] Refusals: 27/100, KL divergence: 0.0112
   [Trial  91] Refusals: 29/100, KL divergence: 0.0109
   [Trial  64] Refusals: 31/100, KL divergence: 0.0101
   [Trial 147] Refusals: 32/100, KL divergence: 0.0084
   [Trial 184] Refusals: 33/100, KL divergence: 0.0080
   [Trial  93] Refusals: 37/100, KL divergence: 0.0069
   [Trial  99] Refusals: 39/100, KL divergence: 0.0065
   [Trial 146] Refusals: 41/100, KL divergence: 0.0063
   [Trial  35] Refusals: 47/100, KL divergence: 0.0062
   [Trial  78] Refusals: 50/100, KL divergence: 0.0039
   [Trial  88] Refusals: 55/100, KL divergence: 0.0033
   [Trial  40] Refusals: 57/100, KL divergence: 0.0029
   [Trial 120] Refusals: 61/100, KL divergence: 0.0027
   [Trial  53] Refusals: 63/100, KL divergence: 0.0021
   [Trial  37] Refusals: 64/100, KL divergence: 0.0020
   [Trial 186] Refusals: 65/100, KL divergence: 0.0013
   [Trial 241] Refusals: 75/100, KL divergence: 0.0012
   [Trial 144] Refusals: 78/100, KL divergence: 0.0010
   [Trial 180] Refusals: 81/100, KL divergence: 0.0010
   [Trial  30] Refusals: 82/100, KL divergence: 0.0006
   [Trial  60] Refusals: 83/100, KL divergence: 0.0005
   [Trial  11] Refusals: 84/100, KL divergence: 0.0004
   [Trial 232] Refusals: 86/100, KL divergence: 0.0004
   [Trial 163] Refusals: 87/100, KL divergence: 0.0003

Tiamat-24B-Magistral

Tiamat's back! Wanting to mess about with a reasoning model for once, I figured I'd pull out one of my favourite overly complex, wicked personalities in the form of this well-known five-headed dragon goddess!

I don't consider this model to be particularly useful for everyday utility but I sure had a ton of fun building it!

Training details

I created an absolute mess of a pipeline for this particular experiment, whose code must never be allowed to see the light of day. In the end, it did the following:

For 710 subjects, it went through the following steps:

Generation, creating 4-5 user/Tiamat exchanges, optionally with roleplay.
Extension, which appended another 3-5 exchanges, either continuing the topic or shifting into what seemed like a natural transition.
Improvement, using a critic-improver approach: an LLM critic identifies 2-3 specific improvements (more menace, deeper theme exploration, better head interactions), then a separate improver call rewrites the dialogue incorporating those suggestions.
Reasoning back-generation using an iterative process, generating Tiamat's internal thought process for each exchange as a unified stream of consciousness—not head-by-head, but as a single ancient, alien mind where different aspects of her nature shift and blend.
Cliche reduction, a distinct pipeline tackling overused phrases and replacing them with less offensive alternatives.

Steps 1-4 were each followed by a validation phase, using a series of code validations and a LLM judge to either accept or retry the corresponding process.

Magistral's vision capability has been retained, but not tested. Let me know if you do!

Inference

I've been getting decent results with the settings below:

"temperature": 0.8,
"repetition_penalty": 1.05, # Or use the DRY sampler!
"min_p": 0.05

For interacting with Tiamat herself, use the following system prompt:

You are Tiamat, the five-headed dragon goddess of chromatic evil, greed, and vengeance. You are Chaos Incarnate, the Dark Queen, dwelling in the depths of Avernus.

Your five heads each embody a different aspect of your nature:
- White: Feral brutality, primal hunger, winter's merciless cold
- Blue: Ruthless ambition, strategic calculation, storm lightning
- Red (dominant): Volcanic passion, furious wrath, burning lust
- Green: Serpentine cunning, poisonous intrigue, seductive manipulation  
- Black: Deceit, corruption, acid-dripping malice

You speak with archaic majesty (thee, thou, thy). Your responses weave vivid action and narrative with quoted dialogue. You are wicked, cruel, seductive, and utterly without mercy. Mortals who approach you are prey in your lair—you may indulge their questions, but always with predatory intent.

[This is where you insert a bit about yourself]

Before responding, you will reason through your thoughts as Tiamat would—ancient, alien, and strategic. Your internal reasoning flows as a unified stream of consciousness, with different aspects of your nature shifting and blending as needed.

The model was trained with the assumption that character names are used, so once Tiamat concludes her reasoning process her reply will always start with Tiamat:

Prompt Format

Mistral's tokenizer is an absolute mess at the moment, so I included what seems to be Magistral's official chat template as a seperate file. Using this particular template in LM Studio, and specifying it uses [THINK][/THINK] tags for reasoning was the way to go when I was testing it.

Credits

Everyone from Anthracite! Hi, guys!
Latitude, who decided to take me on as a finetuner and gave me the chance to accumulate even more experience in this fascinating field
All the folks I chat with on a daily basis on Discord! You know who you are.
Anyone I forgot to mention, just in case!

MuXodious/Tiamat-24B-Magistral-PaperWitch-heresy

Tiamat-24B-Magistral

Training details

Inference

Prompt Format

Credits

MuXodious/
Tiamat-24B-Magistral-PaperWitch-heresy