---
base_model: HuggingFaceTB/SmolLM2-135M-Instruct
library_name: peft
pipeline_tag: text-generation
language:
  - es
tags:
  - lora
  - sft
  - spanish
  - agent
  - technical-assistant
license: apache-2.0
---

# Omegus

Omegus is a Spanish technical chatbot model package for the Charlie / Omega agent architecture.

It is designed to respond as a precise technical assistant with progressive status reporting, software architecture judgment, and clear explanations of the Omega framework.

This repository currently contains a first local demo LoRA adapter. It is intentionally small and should be treated as a prototype checkpoint, not a production-quality assistant yet.

Source material:

- `../Charlie-Skill.md`
- `../spec_maestra_framework_unificado_v0.3.md`

The published adapter fine-tunes `HuggingFaceTB/SmolLM2-135M-Instruct` with LoRA using a compact chat dataset in `data/charlie_omega_sft.jsonl`.

## Training Run

First published adapter:

- Base model: `HuggingFaceTB/SmolLM2-135M-Instruct`
- Method: LoRA SFT
- Local hardware: Apple Silicon MPS
- Dataset size: 20 chat examples
- Epochs: 1
- Train loss: 3.502
- Eval loss: 3.588

Hugging Face Jobs training was attempted, but the account did not have enough prepaid credit balance at the time. The current adapter was trained locally instead.

## Intended Use

- Spanish technical chatbot
- Software architecture and code-review assistant behavior
- Omega framework explanation and synthesis
- Prototype agent-persona research

## Local Dry Run

From this folder:

```bash
uv run train_sft.py
```

This trains locally if your machine has the needed compute.

## Push To Hugging Face

The default Hub target is `TOKETTER/Omegus`. With a logged-in Hugging Face session:

```bash
export HUB_MODEL_ID="TOKETTER/Omegus"
uv run train_sft.py
```

The script pushes LoRA adapter/checkpoints to the Hub when `HUB_MODEL_ID` is set.

## Quick Load

```python
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

repo = "TOKETTER/Omegus"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoPeftModelForCausalLM.from_pretrained(repo)
```

## Recommended Cloud Job Shape

Default Hub target:

```bash
TOKETTER/Omegus
```

For a cheap demo on Hugging Face Jobs:

- Flavor: `t4-small` or similar low-cost GPU
- Timeout: `1h`
- Base model: `HuggingFaceTB/SmolLM2-135M-Instruct`

For a better small assistant:

- Flavor: `a10g-large`
- Timeout: `2h`
- Increase dataset size before training

## Next Dataset Upgrade

The included dataset is intentionally small so the training pipeline is easy to inspect. The next quality step is to expand it into 200-500 instruction examples extracted from the two source docs, with separate examples for:

- Charlie activation and progressive logging
- Code review and bug triage behavior
- Ω framework explanations
- Ω6 functional consciousness caveats
- Mathematical definitions and architecture summaries