VERO Mini V3
VERO Mini V3 is an independent assistant created by Angelo Fontela at Magma Studios. It is based on Qwen/Qwen2.5-1.5B-Instruct and adapted with a small, curated instruction set for VERO identity, Spanish/English behavior, coding, mathematics, safety, uncertainty, structured output, and everyday questions.
What changed from VERO-Mini-V2
- Base model increased from approximately 0.5B to approximately 1.54B parameters.
- The base is a general instruction model rather than only a small coder model.
- The VERO adapter was trained with response-focused loss on 159 training examples and 11 validation examples.
- The release includes a default chat template with VERO identity and behavior guidance.
Evaluation snapshot
On a held-out 20-prompt regression suite, the candidate correctly handled VERO identity, Spanish answers such as Tokio, exact listo output, the Python out-of-range index in items[len(items)], basic coding, mathematics, safety, and uncertainty. It can still be verbose and may wrap JSON-only responses in Markdown fences; callers that require raw JSON should validate and normalize the output.
This evaluation is a small engineering regression suite, not a claim of broad human-level intelligence.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "MagmaStudios/VERO-Mini-V3"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
messages = [{"role": "user", "content": "Escribe una función de Python que sume una lista."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
Limitations
VERO Mini V3 can make factual, coding, language, and formatting mistakes. Verify generated code before execution. Do not use it as the sole basis for medical, legal, financial, security-critical, or other high-impact decisions. It does not know live web information unless an application connects it to a search or retrieval tool. Never put private API keys in prompts or source code.
Attribution
The base model is Qwen2.5-1.5B-Instruct, released under Apache 2.0. The training mixture includes original Magma Studios examples, the public MagmaStudios/VERO-Mini-Training dataset, and replay examples derived from bigcode/self-oss-instruct-sc2-exec-filter-50k as documented by the previous VERO release. Respect the licenses and attribution requirements of each source.