olob0/finevision-mini-thinking

FineVision-mini Thinking

FineVision-mini is a slice I made of HuggingFaceM4/FineVision: 169 of its image subsets, 101,321 rows (~40 GB) out of FineVision's 24.2M rows / 4.65 TB (about 0.4% of the rows, 0.9% of the bytes), sampled with a fixed seed. The 16 text-only subsets were left out.

This dataset is that slice, fully translated and augmented with reasoning, published in increments: each batch processes more rows of FineVision-mini and is appended here, until the whole slice is covered.

For every row:

• —the entire conversation is translated into a single target language, questions and answers alike. The images are the only thing that stays as is. Code, identifiers, URLs and text visible inside code are kept verbatim as part of the translation rules.
• —every assistant turn gets a reasoning trace in the same language, explaining how the answer follows from the image and the question.

What is different

Everything else (images as original bytes, source, the eight *_ratings / *_min columns) is unchanged.

The source conversation is not kept in this dataset. Rows tagged en-US are the model's English rendering of the source and are usually identical to it, but they are still generated text, not a verbatim copy. The untranslated conversations are the texts column of FineVision.

The reasoning is written so that image + question -> reasoning -> response: it describes the visual and logical evidence, never the answer itself, and never implementation details found only in the response (for image-to-code rows it describes layout, colors, typography. Not the code).

Languages

Each row is in one language: a row in pt-BR has its whole conversation and reasoning in Portuguese, and no other version of it exists in the dataset.

The 21 languages are mixed across rows, split evenly inside every subset (so every subset covers every language), and the language column tells which one a row is in.

Filter on it to get a single-language subset: vi-VN zh-CN en-US th-TH id-ID ms-MY fil-PH ja-JP km-KH lo-LA my-MM ta-IN tet-TL ko-KR pt-BR fr-FR de-DE ru-RU es-ES ar-001 jv-ID

ds.filter(lambda r: r["language"] == "pt-BR")

<img src="assets/languages.png" alt="Rows per language" width="400">

Generation

• —Model: Gemma 4 12B IT (QAT, UD-Q4_K_XL), mmproj-F32, MTP draft model for speculative decoding, via llama-server (llama.cpp).
• —Sampling: temperature 1.0, top-p 0.95, top-k 64, reasoning mode off, --image-max-tokens 1024.
• —Prompt: one request per row with all images and all question/response pairs. A fixed system prompt defines the causal rule above, the translation rules (translate the prose, keep code, identifiers, URLs and visible text inside code verbatim) and the output format. The target language is passed as tag (English name), e.g. lo-LA (Lao), because the bare tag was misread for rare languages.

Rules and validation

• —Eligible rows: 1-4 images and 1–30 question/response pairs (the model loses count above that).
• —Output must be valid JSON with exactly one item per pair. All fields strings. r may be empty only when the source response was empty.
• —Non-Latin target languages (th, lo, km, my, ta, ko, ja, zh, ru, ar) are checked by Unicode script. A reasoning trace in the wrong script is rejected. Latin-script languages cannot be checked this way.
• —Generation is capped at roughly twice the expected output length. Truncated or looping outputs are rejected.
• —Up to 3 attempts per row, then a second pass with lower concurrency. Rows the model still fails on (about 0.2%) are omitted, nothing is patched by hand.

License and credits

Images and source conversations come from FineVision and keep the license of their original subset: see the FineVision dataset card for per-source terms. Translations and reasoning traces are model-generated with Gemma 4 and are subject to the Gemma Terms of Use.

Credits go to HuggingFaceM4 for FineVision and the source datasets it aggregates, to Google for Gemma, to unsloth for the GGUF quantizations, and to the llama.cpp project for the inference server.