Huvm

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

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

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.