Self-reported ViDoRe V3 results (8 public retrieval tasks) for Verm1ion/ColTurk-VDR-Qwen3VL-4B-v1.0 — mean NDCG@10 = 0.5584 (full corpus, all queries, MaxSim). ModelMeta PR: embeddings-benchmark/mteb#4796.

Files: results/Verm1ion__ColTurk-VDR-Qwen3VL-4B-v1.0/d56c7bbc278ba2fe4ac1c255fb0e55dd46b40bad/ — 8 task JSONs (live dataset_revisions, mteb_version 2.15.3) + model_meta.json.

Checklist

• My model has a model sheet, report, or similar — https://huggingface.co/Verm1ion/ColTurk-VDR-Qwen3VL-4B-v1.0
• My model has a reference implementation in mteb/models/model_implementations/, this can be as an API. Instruction on how to add a model can be found here
  • No, but there is an existing PR model: add Verm1ion/ColTurk-VDR-Qwen3VL-4B-v1.0 mteb#4796
• The results submitted are obtained using the reference implementation
  • Note: results were produced with the published, self-contained eval harness (identical model-loading path + MaxSim scoring over the full corpora, all queries; seeded bootstrap 95% CIs): https://github.com/Verm1lion/ColTurk-VDR/blob/main/scripts/eval/eval_colturk_checkpoint.py — raw JSONs in that repo. Happy to re-run via mteb.evaluate and update if preferred.
• My model is available, either as a publicly accessible API or publicly on e.g., Huggingface
• I solemnly swear that for all results submitted I have not trained on the evaluation dataset including training splits. If I have, I have disclosed it clearly.
  • Training data (ColPali train set) vs. ViDoRe V3 contamination was additionally checked empirically: perceptual-hash scan over training images x all 8 V3 corpora found 0 exact duplicates (report: https://github.com/Verm1lion/ColTurk-VDR/blob/main/STAGE1_VALIDITY_REPORT.md).