lucas-mella
/

Daimon-R

@@ -1,70 +1,87 @@
----
-license: apache-2.0
-base_model: Qwen/Qwen3-4B-Instruct-2507
-tags:
-  - gguf
-  - daimon
-  - local-assistant
-  - lora
-language:
-  - es
-  - en
----
-# Daimon-R 🤖
-![Daimon-R](daimon-R.png)
-The flagship local brain of **Daimon** — reasoning + coding. Runs on a GPU via llama.cpp.
-**Daimon-R is the flagship brain of Daimon** — a local-first personal AI
-assistant. Daimon-R = `Qwen/Qwen3-4B-Instruct-2507` (quantized GGUF) + a Daimon LoRA, served
-locally with llama.cpp. The weights are open — use them standalone today.
-## What the Daimon LoRA adds
-The LoRA tunes **persona and behavior**, not raw coding (that stays at the base's level):
-- A consistent **Daimon identity** (first person, local-first assistant) instead of a generic
-  "language model" voice.
-- **Canvas convention**: when you ask it to build a web/app/UI it replies with a single
-  self-contained ```html``` block that renders live — no preamble, no copy-paste.
-- Concise **Rioplatense Spanish** for chat, plus a voice-friendly register for TTS.
-## Benchmarks (real, measured locally)
-Daimon uses this model to pick JSON tool-call actions for its browser co-pilot, computer-use,
-and code-editing agents (e.g. `{"action":"click","ref":3}`). Measured on 198 held-out
-synthetic scenarios — disjoint vocabulary/phrasing from training, zero string-level overlap
-with the training set, graded objectively (not by text similarity):
-| Metric | Base Qwen3-4B + identity-only LoRA | **Daimon-R** |
-|---|---|---|
-| Valid JSON | 97.5% | **100.0%** |
-| Correct action chosen | 65.8% | **94.4%** |
-| Correct action + correct fields (ref/path/command/...) | 39.2% | **94.4%** |
-Persona/behavior suite (identity, canvas-HTML convention, Rioplatense style, no-refusal): **12/12**, no regression from the identity-only checkpoint.
-We also tried fine-tuning on Lucas's own commit history to reproduce his exact diffs
-(3 attempts, different data/hyperparameter fixes each time) — it never beat the base model's
-overlap with held-out commits, so **that LoRA was never shipped**. We only publish results
-that actually win a real, pre-registered comparison.
-## Run standalone
-```bash
-huggingface-cli download lucas-mella/Daimon-R --local-dir ./daimon-models
-# llama.cpp: load the base GGUF and apply the Daimon LoRA
-llama-server -m ./daimon-models/Qwen3-4B-Instruct-2507-Q4_K_M.gguf \
-  --lora ./daimon-models/daimon-r-lora-f16.gguf -c 8192
-```
-Exposes an OpenAI-compatible endpoint (default http://localhost:8080/v1).
-## Files
-- `Qwen3-4B-Instruct-2507-Q4_K_M.gguf`
-- `daimon-r-lora-f16.gguf`
-## Daimon — coming soon
-These weights are the brain of **Daimon**, a local-first personal assistant that runs on
-your own machine: real-time local voice, a co-pilot browser that Daimon and you share, a
-canvas for apps & prototypes, and hybrid local/cloud routing. The full app isn't public
-yet — these open models are a complement you can already build on. Watch
-[@lucas-mella](https://huggingface.co/lucas-mella) for the release.

+---
+license: apache-2.0
+base_model: Qwen/Qwen3-4B-Instruct-2507
+language:
+- es
+- en
+tags:
+- daimon
+- lora
+- tool-calling
+- code
+- spanish
+library_name: gguf
+---
+<div align="center">
+<img src="https://huggingface.co/lucas-mella/Daimon-R/resolve/main/daimon-R.png" width="120" alt="Daimon"/>
+<br/><br/>
+<img src="https://huggingface.co/lucas-mella/Daimon-R/resolve/main/daimon-palabra.png" width="160" alt="daimon"/> <b style="font-size:1.6em;vertical-align:middle">r</b>
+</div>
+---
+## 🇪🇸 Español
+**daimon r** es el modelo **razonador de 4B** (fine-tune LoRA de `Qwen3-4B-Instruct-2507`), el hermano mayor de `daimon x`: mismo asistente de **código, agente (tool-calling) y conversación**, con foco en **español nativo**, pero con más capacidad de razonamiento.
+> Forma parte del proyecto **Daimon** — *próximamente / coming soon*.
+### 🧠 Con qué se entrenó
+Fine-tune **LoRA (QLoRA 4-bit)** sobre el mismo mix curado que `daimon x`, 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 |
+|---|---|---|---|
+| Qwen3-4B-Instruct-2507 | 4.979 | 0.762 | **0.633** |
+`eval_loss < train_loss`, y notablemente más bajo que `daimon x` (0.633 vs 0.868) → el 4B aprende mejor.
+### ⬇️ Descargar (opcional — solo si querés usarlo)
+```bash
+huggingface-cli download lucas-mella/Daimon-R
+# correr con llama.cpp (base GGUF + adapter LoRA):
+llama-server --model Qwen3-4B-Instruct-2507-Q4_K_M.gguf \
+             --lora daimon-r-lora-f16.gguf --alias daimon-r
+```
+---
+## 🇬🇧 English
+**daimon r** is the **4B reasoning** model (LoRA fine-tune of `Qwen3-4B-Instruct-2507`), the big sibling of `daimon x`: the same **code, agentic tool-calling and conversation** assistant with a focus on **native Spanish**, but with stronger reasoning.
+> Part of the **Daimon** project — *coming soon*.
+### 🧠 Training
+**LoRA (4-bit QLoRA)** fine-tune on the same curated mix as `daimon x`, 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 |
+|---|---|---|---|
+| Qwen3-4B-Instruct-2507 | 4,979 | 0.762 | **0.633** |
+`eval_loss < train_loss`, and notably lower than `daimon x` (0.633 vs 0.868) → the 4B learns better.
+### ⬇️ Download (optional)
+```bash
+huggingface-cli download lucas-mella/Daimon-R
+# run with llama.cpp (base GGUF + LoRA adapter):
+llama-server --model Qwen3-4B-Instruct-2507-Q4_K_M.gguf \
+             --lora daimon-r-lora-f16.gguf --alias daimon-r
+```
+---
+<sub>Base: Qwen3-4B-Instruct-2507 (Apache-2.0). LoRA + mix: proyecto Daimon.</sub>