lucas-mella/Daimon-X

  x

---

🇪🇸 Español

daimon x es un modelo de 1.5B (fine-tune LoRA de Qwen2.5-Coder-1.5B-Instruct), pensado como asistente local de código, agente (tool-calling) y conversación, con foco especial en español nativo.

Forma parte del proyecto Daimon — próximamente / coming soon.

🧠 Con qué se entrenó

Fine-tune LoRA (16-bit) sobre un mix curado y balanceado, pesado hacia el núcleo de Daimon (tool-calling / agente):

• Tool-calling / agente — xLAM, ToolACE, Hermes-FC, Toucan (trayectorias multi-turno reales de +495 servidores MCP)
• Código — evol-codealpaca, OpenCodeReasoning, glaive
• Razonamiento — OpenThoughts, OpenR1-Math (con respuestas verificadas)
• Español nativo — BSC-LT m-personas, projecte-aina MentorES, aya (nativo, no traducido)
• In-house — acciones de agente + identidad de Daimon

📊 Benchmarks

Base	Ejemplos	train_loss	eval_loss
Qwen2.5-Coder-1.5B-Instruct	4.967	0.916	0.868

eval_loss < train_loss → aprendió sin sobreajustar.

💻 Requisitos de hardware

Recurso	Mínimo	Recomendado
Disco (GGUF q8_0 + LoRA)	~2 GB	~2 GB
RAM	4 GB	8 GB
GPU	no hace falta (corre en CPU)	≥ 2 GB VRAM (offload completo)

Modelo muy liviano: corre fluido en CPU en casi cualquier PC o notebook.

⬇️ Descargar (opcional — solo si querés usarlo)

huggingface-cli download lucas-mella/Daimon-X
# correr con llama.cpp (base GGUF + adapter LoRA):
llama-server --model qwen2.5-coder-1.5b-instruct-q8_0.gguf \
             --lora daimon-sft-lora-f16.gguf --alias daimon-x

---

🇬🇧 English

daimon x is a 1.5B model (LoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct) — a local assistant for code, agentic tool-calling and conversation, with a strong focus on native Spanish.

Part of the Daimon project — coming soon.

🧠 Training

LoRA (16-bit) fine-tune on a curated mix, weighted toward Daimon's tool-calling/agent core:

• Tool-calling / agent — xLAM, ToolACE, Hermes-FC, Toucan (real multi-turn trajectories from 495+ MCP servers)
• Code — evol-codealpaca, OpenCodeReasoning, glaive
• Reasoning — OpenThoughts, OpenR1-Math (verified answers)
• Native Spanish — BSC-LT m-personas, projecte-aina MentorES, aya (native, not translated)
• In-house — Daimon agent-actions + identity

📊 Benchmarks

Base	Examples	train_loss	eval_loss
Qwen2.5-Coder-1.5B-Instruct	4,967	0.916	0.868

eval_loss < train_loss → learned without overfitting.

💻 Hardware requirements

Resource	Minimum	Recommended
Disk (q8_0 GGUF + LoRA)	~2 GB	~2 GB
RAM	4 GB	8 GB
GPU	none (CPU works)	≥ 2 GB VRAM (full offload)

Very lightweight — runs smoothly on CPU on almost any PC or laptop.

⬇️ Download (optional)

huggingface-cli download lucas-mella/Daimon-X
# run with llama.cpp (base GGUF + LoRA adapter):
llama-server --model qwen2.5-coder-1.5b-instruct-q8_0.gguf \
             --lora daimon-sft-lora-f16.gguf --alias daimon-x

---

_{Base: Qwen2.5-Coder-1.5B-Instruct (Apache-2.0). LoRA + mix: proyecto Daimon.}