logo-mistral-q4_k_m

🧠 Logo-Mistral Q4_K_M

Mistral Small 24B model fine-tuned on the LLT (Logic Layer Tool) system — quantized Q4_K_M

---

📌 Description

Logo-Mistral is a language model based on Mistral Small 24B Instruct, fine-tuned to apply the formal reasoning system LLT (Logic Layer Tool), a framework Structured axiomatic model that separates logical-deductive reasoning from heuristic-creative reasoning.

The model is distributed in GGUF quantized Q4_K_M format, optimized for local execution using tools such as llama.cpp, Ollama, or LM Studio.

---

🔬 The LLT System

LLT is a structured reasoning framework that operates in two main modes:

LLT Output Structure

Each response follows a deterministic pipeline format:

[MODE: V-LOGIC / H-LOGIC]
[INGEST: <input parsing>]
[PARSE: <problem structure>]
[TRAP-SCAN: <cognitive trap detection>]
[EXECUTE: <axiom or lemma applied>]
[VERIFY: <consistency check>]
[ARTICULATE: <final response with confidence>]

📊 Training Dataset

The model was fine-tuned on a structured dataset of over 7,000 examples (instruction/output pairs) built around the LLT system. The dataset covers:

• Formal Logic — syllogisms, implications, contradictions, paradoxes
• Heuristic Problem Solving — classic puzzles, lateral thinking, brain teasers
• Probability and Combinatorics — dependent/independent events, complement, geometric distributions
• Philosophy of Mind and Epistemology — knowledge vs. belief, truth, ontological coherence
• Practical Reasoning — cause-and-effect analysis, functions, structural constraints
• Classic Puzzles — river crossing, weighing, torches on the bridge, ropes, and time

The dataset is bilingual (Italian / English).

---

🚀 How to use it

With llama.cpp

./llama-cli \
-m logo-mistral-q4_k_m.gguf \
-p "Use the LLT system: If it rains, the ground is wet. The ground is not wet. What do you conclude?" \
-n 512

With Ollama

ollama run CiroN2022/logo-mistral-q4_k_m

With LM Studio

Download the .gguf file and load it directly into the LM Studio interface.

---

💬 Usage Examples

Example 1 — Modus Tollens (V-LOGIC)

Input:

Use the LLT system: If it rains, the ground is wet. The ground is not wet. What do you conclude?


**Output:**

[MODE: V-LOGIC]
[INGEST: P1: Raining→Wet, P2: ¬ Wet]
[PARSE: PROP: P→Q, PROP: ¬ Q]
[TRAP-SCAN: No trap detected]
[EXECUTE: Apply A21 [P→Q, ¬ Q ⊢ ¬ P] (MODUS_TOLLENS)]
[VERIFY: A22 (NON_CONTRADICTION) satisfied]
[ARTICULATE: I conclude that it is not raining (¬ P). If the implication is true and the consequent is false, the antecedent must be false. Confidence: 1.0]

---

Example 2 — Lateral Thinking (H-LOGIC)

Input:

Use LLT: An electric train is traveling north at 100 km/h. The wind is blowing from the south at 50 km/h. Which way is the smoke going?

Output:

[MODE: H-LOGIC]
[INGEST: TRAIN: Electric, DIR: North, WIND: South]
[PARSE: DEFAULT: Trains_produce_smoke]
[TRAP-SCAN: T20 (Literal Overshoot)]
[EXECUTE: A42 [REVERSE→INSIGHT] - Reverse: 'Does an electric train produce smoke?']
[VERIFY: Electric trains = no combustion]
[ARTICULATE: The smoke isn't going anywhere. The train is electric. The hidden structure is the implicit assumption. Confidence: 1.0]

---

Example 3 — Probabiity

Input:

From a standard deck, you draw two cards without replacement. What is the probability the first is a heart and the second is a spade?

Output:

[MODE: V-LOGIC]
[INGEST: HEARTS: 13, SPADES: 13, TOTAL: 52, DRAWS: 2 without replacement]
[PARSE: DEPENDENT: P(heart first) × P(spade second | heart first)]
[TRAP-SCAN: No trap]
[EXECUTE: L5.45 [DEPENDENT-SUITS] - Multiply conditional probabilities]
[VERIFY: P(heart first) = 13/52 = 1/4. P(spade second | heart first) = 13/51. P = (1/4)×(13/51) = 13/204]
[ARTICULATE]: 13/204 or approximately 6.37%. Confidence: 1.0]

---

⚙️ Technical Specifications

---

⚠️ Limitations

• The model specializes in LLT formal reasoning; It may perform less well on generic, unstructured tasks.
• The Q4_K_M quantization introduces a slight loss of precision compared to the original float16 model.
• Some ambiguous scenarios may generate output completely wrong.

---

📄 License

This model is distributed under the Apache 2.0 license. The Mistral base model is subject to the Mistral AI license terms.

---

👤 Author

Created by CiroN2022 🔗 https://huggingface.co/CiroN2022