Modelpublic

uran1um1/tenorio-1.5b-GGUF

sourceHugging Facemitupdated 22d agoView on Hugging Face

0likes339downloads

Model Card

Tenorio 1.5b

Overview

This is the second and most recent model in the Tenorio family of large language models. This model is based on and is a fine tune of Qwen2.5 1.5b Instruct. It is based on a highly similar and nearly identical tuning architecture as is seen on Tenorio 0.6b. The purpose of this tune is to improve the Spanish language proficiency and formatting in the model's outputs. I want to note that this is a hobby project and results are not guaranteed.

Observed Improvements

—Better formatting (usage of Markdown, etc) when it comes to answering with the Spanish language. This can be seen in BASE_LICENSE in the root of the repository.
—More inquiries (asking the user for input or preference) in general conversation. This can be seen in LICENSE in the root of the repository.

This model is a fine tune of Alibaba's Qwen2.5 1.5b Instruct, 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.

Licensing

—As this model is a direct fine tune of the Unsloth Qwen2.5 1.5b Instruct, it is bound to the Apache 2.0 license from the base model.
—This specific fine tune (modifications) are licensed under the MIT license.

Training Architecture

Below is an outline on the training process used to fine tune this model.

—Trained on roughly 100+ public domain books in the Spanish language via CPT (continued pretraining).
—Instruction improvement LoRA tuning using 7077 Q&A pairs in the Spanish language distilled from Bonsai 8b using an automated self-prompting distillation script.

The training process is almost identical regarding data between the different models in the Tenorio series.

Benchmarks

Although I have not run any formal standardized benchmarks, I have wrote my own small benchmarking script which makes the base and tuned model compete on a series of 592 questions in the Spanish language. Below is a detailed account on the function and process of the script.

—The two models (Q8_0 quantization with Unsloth) are loaded into the environment using an LMStudio (llmster/lms) backend.
—For each question, each model is prompted the exact same question.
—A stronger judge model (GPT-OSS 20b) judges the output, as phrased with the prompt below:
—Judge Prompt:

    Which answer is better BASED SOLELY ON SPANISH GRAMMATICAL CORRECTNESS AND NATURAL LANGUAGE ABILITY.
      Output the word DAY for answer A, and NIGHT for answer B. NO EXPLANATION, NO FORMATTING, OUTPUT A SINGLE WORD.

          Answer A: {Answer for Tenorio here}
          Answer B: {Answer for Base here}

          Output now. ONLY 1 WORD.

This means that if Tenorio scores exactly 296, this means that it performed and beat the exact same number of questions as the base model. The results of this mini-benchmark are mentioned below.

Benchmark	Score (Tenorio)	Score (Base)
Spanish Language	315/592	277/592

The benchmarks above are sources from the same benchmark run, as the two models are competing, not individually tested.