CLLG/Qwen3VL-8B-GGUF
Quantized GGUF versions of the CLLG Qwen3-VL 8B model, fine-tuned for ancient Greek document understanding and TEI XML encoding.
This model is part of the
Corpus Liberatum Linguae Graecae (CLLG) project, which aims to build a freely accessible, high-quality corpus of ancient Greek texts. It is supported by the
ANR within the
Programme Inria Quadrant (PIQ).
Model Description
This vision-language model is trained to process images of critical editions of ancient Greek (and Latin) texts and produce structured TEI XML output. It handles the complex page layouts typical of scholarly editions, including:
- Main text in polytonic Greek
- Canonical references (section, paragraph, line numbers)
- Titles and headings
- Footnotes and apparatus criticus markers
GGUF Files
You need both the language model file and the multimodal projector to run this model.
| Filename | Quantization | Size | Use case |
|---|
Qwen3VL-8B-synth_real.Q4_K_M.gguf | Q4_K_M | ~5 GB | Recommended — good balance of size and quality |
Qwen3VL-8B-synth_real.Q5_K_M.gguf | Q5_K_M | ~6 GB | Higher quality |
Qwen3VL-8B-synth_real.Q8_0.gguf | Q8_0 | ~9 GB | Near-lossless |
mmproj-BF16.gguf | BF16 | — | Vision projector — required for all variants |
Usage
1llama-cli \
2 --model Qwen3VL-8B-synth_real.Q4_K_M.gguf \
3 --mmproj mmproj-BF16.gguf \
4 --image page.jpg \
5 --prompt "Encode the following page in TEI XML."
Training Data
Fine-tuned on synthetic page images generated by the CLLG pipeline, covering approximately 175,000 Greek pages and 10,000 Latin pages drawn from 4,582 works, with over 5,000 typographic style combinations, complemented by real annotated document pages (synth_real suffix).
Intended Uses
- Automatic TEI XML encoding of ancient Greek critical editions
- Layout analysis and canonical reference detection in scholarly documents
- Research in digital philology and computational humanities
Out of Scope
- Modern Greek text
- Non-document image understanding tasks
- Apparatus criticus and complex critical apparatus (current focus is prose text)
- Poetry and Drama
Project & Funding
This model is developed as part of the CLLG project, funded by the ANR within the PIQ initiative. Institutional partners include Persée and Biblissima.
Citation
If you use this model in your research, please cite the CLLG project and acknowledge ANR/PIQ funding.
License
Apache 2.0 — see
LICENSE.
