Huvm

Huvm is a compact, instruction-tuned language model fine-tuned from Qwen/Qwen2.5-1.5B-Instruct using LoRA. It is designed to be fast, precise, and carry a distinct personality — a blend of 50% Grok (direct, lightly sarcastic, no-nonsense) and 50% Claude (articulate, deep, helpful).

The name comes from "humm" (the sound of thinking) plus the letter V for speed and truth.

Model Description

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Fine-tuning method: LoRA (r=16, alpha=32) via TRL SFTTrainer
• Training: CPU (float32), 5 epochs, ~40 curated examples
• Inference formats: Hugging Face Transformers, GGUF (quantized for llama.cpp / mobile)
• Dataset: Custom dataset

Intended Use

Huvm is built for local, offline use — on a desktop CPU or a mobile device via GGUF. It handles:

• 💻 Code — explains what it does and why
• ➗ Math — shows the reasoning, not just the answer
• 🎨 Creativity — avoids generic responses
• 💬 Multilingual chat — replies in the user's language

The model is small enough to run on phones with ~6 GB RAM and obtains reasonable quality given its size. It is not intended for high-stakes or safety-critical applications.

Personality & Tone

Trait	Behavior
Directness	Gets to the point, minimal filler
Sarcasm	Light and playful, never insulting
Depth	Explains complex topics simply when asked
Honesty	Refuses to invent facts, links, or stats
Language	Responds in the user's language (pt/en/es, etc.)

Bias, Risks & Limitations

• Small model — 1.5B parameters; may hallucinate or repeat generic phrasing on complex topics.
• Limited knowledge cutoff — reference date of 2026-08-30; not updated in real time.
• Fine-tuned on a small dataset — identity and behavior are reinforced, but coverage is narrow.
• No safety tuning beyond the base model — use with judgement.

Training Details

• Data: A custom JSONL dataset containing behavior examples (identity, language, code, math, creativity, knowledge) and anti-patterns (what not to do).
• Technique: LoRA applied to the attention projection layers.
• Hardware: AMD Ryzen 5 + 16 GB RAM, CPU-only (AMD GPU lacks Windows ROCm support).

Quantization

A q4_k_m GGUF file (~940 MB) is provided for llama.cpp and mobile apps:

llama-cli -m huvm-q4_k_m.gguf -p "Quem e voce?"

Usage (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("your-username/huvm")
tokenizer = AutoTokenizer.from_pretrained("your-username/huvm")

messages = [
    {"role": "system", "content": "Eu sou o Huvm, um assistente de IA..."},
    {"role": "user", "content": "Faca uma funcao em Python que inverte uma string"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Acknowledgements

Built on Qwen 2.5 (Alibaba), TRL (Hugging Face), and llama.cpp.