Qwen3.5-0.8B (base) vs. its version fine-tuned for LaTeX OCR comparison

Description

Models in this comparison:

• Base Model: unsloth/Qwen3.5-0.8B (Q8_0-GGUF used in this comparison)
• Fine-tuned model: darkhan-awsjp202305/Qwen3.5-0.8B-latex-ocr-v0.1 (Q8_0-GGUF used in this comparison)

Training Dataset:

• unsloth/LaTeX_OCR

Legend:

• Source Image - The image that was given to the model with a request to produce LaTeX output. These images are from the unsloth/LaTeX_OCR dataset. However, the 80 images in this table are from the "test" slice which was not part of the LaTeX OCR model's fine-tuning training data.
• Result - Model output (LaTeX) rendered with the KaTeX library.
• Completed (retries) - yes / no - whether the model produced output or went into an infinite loop, if the model failed to produce a bound result at first, the generation was retired up to 5 times.