Tenorio 0.6b
Overview
Tenorio 0.6b is a edge/tiny model tuned to have increased proficiency in the modern Spanish language. It has been trained to respond without reasoning and have better formatting and grammatical correctness when answering in the Spanish language. GGUF version here.
Training Architecture
Below is an outline on the training process used to fine tune this model.
- Trained on roughly 100+ (not verified exactly) public domain books in the Spanish language via CPT (continued pretraining).
- Instruction improvement tuning using 7077 Q&A pairs in the Spanish language generated by Bonsai 8b.
This model is a fine tune of Alibaba's Qwen3 0.6b, which itself is released under the Apache 2.0 license. Any issues regarding the licensing of this tune can be communicated to the author of this tune directly on HuggingFace and will be addressed.
The license for the base model can be found in the BASE_LICENSE.md file in the repository.
Observed Improvements
- Better context awareness and less hallucination in Spanish.
- Near complete destruction of the reasoning ability (this was done on purpose to reduce hallucinations and infinite loops).
- More concise and natural answers.
Benchmarks
As this model is a hobby project, I have not had much time or patience to benchmark it professionally. That is why I wrote a tiny custom benchmarking script to test the effectiveness of the model. Below is an outline of how the benchmarking worked, step by step.
- Using an LMStudio backend, the two Q8_0 Quantized versions of the base and tuned model are loaded.
- Over a series of 80 questions (per benchmark), the script asks both of the models the same question.
- A stronger model (in this case, GPT-OSS 20b) receives a basic instruction prompt asking it to judge between the two models. In my script, it was asked to judge it based on the grammar and quality of the language. The benchmarks are only for measuring the quality of the Spanish language output, not the individual quality or accuracy of the content.
- The points are tracked. In this case, scoring a 40/80 means that the fine tuned model beat the base model in exactly half of the questions.
| Benchmark | Score |
|---|---|
| Creative writing | 40/80 |
| General STEM | 42/80 |
| Coding & CS | 45/80 |
I would like to note that as these benchmarks are quite basic, I would appreciate and accept open benchmarks from anyone in the community willing to test this out.