Trenuje modele z pasji i dla społeczności. Jeśli moja praca ułatwia Ci życie lub oszczędza czas, każda postawiona kawa pomaga w opłaceniu kolejnych treningów GPU! ☕👇

Buy Me A Coffee

Model Description

Leonidas-4B is a fine-tuned Polish reasoning model built on the Qwen3.5-4B hybrid architecture (Mamba + Attention). It was trained using LoRA fp16 on a curated 48k Polish Chain-of-Thought dataset, with native <think> reasoning blocks.

Forged by Hattori Hanzo — because an idiot admires complexity, a genius admires simplicity.

"An idiot admires complexity, a genius admires simplicity." — Terry A. Davis, TempleOS

Architecture

Property	Value
Base Model	Qwen3.5-4B (hybrid Mamba+Attention)
Parameters	~4B
Training Method	LoRA fp16 (r=16, alpha=16)
Trainable params	0.47%
Training Steps	1500
Final Loss	0.4389
Context Length	32768
Language	Polish 🇵🇱 + English

Training

Dataset: 48k Polish CoT (Chain-of-Thought) — mixed reasoning, math, logic, science
Format: ChatML with native <think> blocks
Platform: Kaggle T4 16GB
Phases: | Phase | Steps | LR | Scheduler | |---|---|---|---| | 1 | 2000 | 2e-4 | linear | | 2 | 3000 | 5e-5 | cosine | | 3 | 5000 | 3e-5 | cosine | | 4 | 5000 | 1e-5 | cosine |

Available Files

File	Size	Description
`leonidas_f16.gguf`	8.41 GB	Full precision fp16
`leonidas_Q8_0.gguf`	4.47 GB	Q8_0 — best quality
`leonidas_Q6_K.gguf`	3.45 GB	Q6_K — great quality
`leonidas_Q4_K_M.gguf`	2.70 GB	Q4_K_M — recommended for most setups

Usage

llama.cpp

bash

./llama-cli \
  -m leonidas_Q4_K_M.gguf \
  -p "Jesteś asystentem AI. Myśl krok po kroku." \
  --chat-template chatml \
  -n 1024

Ollama / OpenWebUI

Compatible with any OpenAI-compatible frontend supporting GGUF + ChatML template.

Recommended Parameters

temperature:        0.6
top_p:              0.95
top_k:              40
repetition_penalty: 1.05

Capabilities

✅ Native Polish reasoning with <think> CoT blocks
✅ Mathematics and logic
✅ Scientific explanations
✅ Sentiment analysis
✅ Code generation (Python, basics)
✅ Quantum physics concepts

Forge Stamp

general.author:   HattoriHanzo1
hanzo.base_model: Qwen3.5-4B
hanzo.leonidas:   Τότε ἐν τῇ σκιᾷ μαχούμεθα
                  ("Then we shall fight in the shade.")

License

Apache 2.0 — base model license from Qwen3.5-4B (Alibaba Cloud).

☕ Wsparcie Projektu

Stworzenie i dopracowanie tego modelu wymagało sporo czasu, testów i zasobów sprzętowych. Jeśli uznajesz go za użyteczny i chcesz wesprzeć dalszy rozwój otwartych projektów, możesz dorzucić swoją cegiełkę:

Buy Me A Coffee

Dzieki za kazde wsparcie, widzimy sie przy kolejnych treningach! 🚀

Romani ite domum 😄 — Qapla'! ⚔️🖖

HattoriHanzo1/Leonidas-4B-GGUF

Leonidas-4B