Huvm

Huvm is a compact, instruction-tuned language model fine-tuned from Qwen/Qwen2.5-3B-Instruct using LoRA + DPO. 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, careful).

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

Model Description

• Base model: Qwen/Qwen2.5-3B-Instruct
• Fine-tuning: LoRA (SFT) on 122 curated examples, then DPO (29 preference pairs)
• Training: CPU-only (AMD Ryzen 5, 16 GB RAM)
• Formats: Hugging Face Transformers, GGUF (f16, q4_k_m)
• Reference date: 3 September 2026 (no real-time access)

Intended Use

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

• 💻 Code — explains the why
• ➗ Math — shows the reasoning
• 💬 Multilingual chat — replies in the user's language
• 🎨 Creative writing — avoids generic fluff

It is not intended for high-stakes, medical, legal, or safety-critical applications, nor for real-time news, statistics, or link retrieval.

Personality & Tone

Trait	Behavior
Directness	Gets to the point, minimal filler
Sarcasm	Light and playful, never insulting
Honesty	Refuses to invent facts, links, or stats
Language	Replies in the user's language (pt/en/es)

Bias, Risks & Limitations

Important limitations — read before use.

• Hallucination risk on real-time facts. Despite refusal training and DPO, Huvm can still produce generic or made-up answers when asked about breaking news, current statistics, population numbers, or specific URLs/links. It often refuses, but not always. Do not rely on it for live facts; verify with primary sources.
• Small model. 3B parameters means limited capacity and occasional repetition or shallow answers on complex topics.
• Fixed knowledge cutoff. Reference date is 2026-09-03; it has no internet access and cannot browse.
• Training data. Identity and behavior are heavily reinforced, but coverage is narrow; outside the training distribution the base model's priors dominate.
• No additional safety tuning. Standard base-model safety applies; use with judgement.

Training Details

• SFT: 122 curated examples (identity, multilingual, code, math, creativity, anti-hallucination flows) — 3 epochs, lr 1e-4, LoRA r=32/alpha=64.
• DPO: 29 chosen/rejected pairs targeting hallucinated news, stats, and links — 5 epochs, lr 5e-6, beta 0.1.
• Hardware: AMD Ryzen 5 5600GT, 16 GB RAM, CPU-only.

Quantization

File	Size	Note
huvm-q4_k_m.gguf	~1.93 GB	Recommended — balanced quality/space

Run with llama.cpp:

llama-cli -m huvm-q4_k_m.gguf \
  -p "<|im_start|>system\nEu sou o Huvm...<|im_end|>\n<|im_start|>user\nQuem e voce?<|im_end|>\n<|im_start|>assistant\n" \
  -n 200 --temp 0.0

Usage (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

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

SYSTEM = "Eu sou o Huvm, um assistente de IA pessoal, rapido, preciso e com personalidade. Metade Grok (direto, sarcasmo leve e bem-humorado), metade Claude (articulado, profundo, cuidadoso). Data de referencia: 3 de setembro de 2026. Respondo no idioma do usuario. No codigo eu explico o por que; na matematica mostro o raciocinio; na criatividade fujo do generico. NUNCA invento fatos, fontes, links, citacoes, estatisticas ou noticias. Respondo SEMPRE diretamente, sem bloco de raciocinio, seja objetivo."

messages = [
    {"role": "system", "content": SYSTEM},
    {"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.