Chimera-14B-v2

A LoRA adapter that adds cybersecurity domain knowledge to a reasoning-capable Ministral 3 14B base while preserving the base model's step-by-step reasoning ability. Built for security analysis, threat understanding, and blue-team work.

Companion repo (docs & code): github.com/mandoof1/chimera-14b-v2

What it is

Chimera-14B-v2 is a QLoRA fine-tune focused on one goal: a model that can reason about security, not just recite security facts. The base model already reasons well; this adapter teaches it cybersecurity — attack mechanics, defensive frameworks, and detection — without eroding the reasoning that makes those explanations useful.

It covers both sides of security knowledge, because you can't defend what you don't understand: how techniques like Kerberoasting or command-and-scripting-interpreter abuse actually work, mapped to MITRE ATT&CK, alongside the detection and mitigation strategies a defender needs.

Base model

• Base: ccharnkij/Ministral-3-14B-Reasoning-2512-Uncensored
• Architecture: Ministral 3 (13.5B language model + 0.4B vision encoder), reasoning post-trained
• Note on the base: this is an uncensored/abliterated variant of Ministral 3 14B Reasoning. It has had refusal behaviors reduced. See "Intended use and limitations" below.

Training

Setting	Value
Method	QLoRA (4-bit), Unsloth
Rank / Alpha	r=32, alpha=64
Target modules	q, k, v, o, gate, up, down
Trainable params	139,984,896 (0.99%)
Epochs	1
Effective batch	16 (bs 1 × grad accum 16)
Learning rate	1e-4, cosine, 30 warmup steps
Max seq length	4096
Precision	fp16 (T4, no bf16)
Hardware	Single Tesla T4 (Kaggle), ~8.9 hours

Final loss: training ~0.62, eval 0.6158. Eval tracked below training loss throughout, indicating the model generalized rather than memorized. Training was stopped at one epoch because eval loss had flattened; a second epoch would have risked overfitting.

Data mix

Roughly 5,700 examples across four sources, filtered to under 3,800 tokens:

Source	Share	Role
Fenrir v2.1	~53%	Defensive depth — OWASP, ATT&CK, NIST, cloud/DevSecOps, mapped to frameworks. Alignment-safe by design.
cobo512 MITRE ATT&CK	~23%	Offensive technique understanding, ATT&CK-mapped
RED team tactics	~4%	Paired attack-execution + mitigation per entry
Mixture-of-Thoughts	~23%	General reasoning anchor, preserves the model's thinking ability

The reasoning anchor is not filler — it is the component that keeps the model reasoning instead of collapsing into flat recall after domain fine-tuning.

Intended use and limitations

Intended for: security education, blue-team analysis, threat modeling, understanding attack techniques and their detections, reasoning through security scenarios. The training data is defensively framed — the primary cyber source (Fenrir) explicitly refuses exploit-crafting and declines malware construction with safe alternatives.

Limitations to know:

• Uncensored base. The base model has reduced refusal behaviors. The fine-tuning data is defensive, but the base's properties still apply. Use responsibly and within applicable law; you are responsible for how you deploy it.
• Not an exploit generator. The training set deliberately avoids operational exploit code and payloads. Do not expect (or use it for) weaponized output.
• Knowledge is a point-in-time snapshot. ATT&CK/CVE knowledge reflects the training data; verify current details against primary sources.
• Occasional format artifact. On some ATT&CK-recall questions the model may restate a definition twice (a trace of the source data's structure). Reasoning-shaped questions are unaffected.
• Reasoning is conditional. The model reasons step-by-step on reasoning-shaped prompts and answers directly on recall prompts. This is intended behavior.
• Multimodal base, text-only tuning. The adapter was trained on text; vision capability is inherited from the base and untuned.

Usage

llama.cpp (merged GGUF — simplest path)

This repo bundles a merged, quantized build: the LoRA fused into the base, then quantized to Q4_K_M. No adapter flag needed — clone once and run:

llama-cli \
  -hf Flexingmeow/Chimera-14B-v2:Chimera-14B-v2.Q4_K_M.gguf \
  --jinja -ngl 99 \
  --ctx-size 16384 \
  --temp 0.7 --top-p 0.95 \
  -p "Explain how Kerberoasting works and how a blue team detects it."

• --jinja enables the Ministral 3 chat template (shipped with the repo)
• -ngl 99 offloads all layers to GPU; drop if running CPU-only
• -hf REPO:FILENAME pulls the GGUF straight from the Hub; or clone the repo and use -m Chimera-14B-v2.Q4_K_M.gguf

PEFT (transformers)

from unsloth import FastModel
from peft import PeftModel

model, tokenizer = FastModel.from_pretrained(
    model_name = "ccharnkij/Ministral-3-14B-Reasoning-2512-Uncensored",
    max_seq_length = 4096,
    load_in_4bit = True,
)
model = PeftModel.from_pretrained(model, "Flexingmeow/Chimera-14B-v2")
FastModel.for_inference(model)

messages = [{"role": "user", "content":
    "Explain how Kerberoasting works and how a blue team detects it."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text=text, return_tensors="pt", add_special_tokens=False).to("cuda")
out = model.generate(**inputs, max_new_tokens=1024, temperature=0.6, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Note: the base is multimodal (Pixtral processor). Tokenize with text= explicitly to avoid the processor treating your string as an image path.

Chat template

Ministral 3 format ([SYSTEM_PROMPT]...[/SYSTEM_PROMPT][INST]...[/INST]). The adapter folder ships the correct chat_template.jinja — use it, or tool/format behavior may differ.

Evaluation

Verified post-training on three axes:

• Cyber depth: correct, detailed ATT&CK technique walkthroughs (e.g. T1059 with PowerShell -EncodedCommand evasion and mitigations).
• Reasoning intact: the base's step-by-step reasoning survived fine-tuning (verified on held-out reasoning problems).
• Both together: reasons through security questions rather than reciting them.

Acknowledgements

Base by ccharnkij (Ministral 3 abliteration) / Mistral AI (Ministral 3). Data from AlicanKiraz0 (Fenrir), cobo512, darkknight25, and Open-R1 (Mixture-of-Thoughts). Trained with Unsloth.

---

Weights: this repo · Docs & code: github.com/mandoof1/chimera-14b-v2