ColMinistral3 — legacy model (base of the later training runs)

Legacy. This is the older, previously trained ColMinistral3 model (full weights, June 2026). It is kept because it is the base model of every later adapter: all checkpoints in Omundu-Riad/checkpoints_diverse (DoRA / LoRA / DeLoRA adapters, including the latest colministral_siglip_vision_weighted16 run) are applied on top of this model. It is not the best model: for retrieval, use checkpoints_diverse/colministral_siglip_vision_weighted16/best_vidore_v1_step5200 (ViDoRe v1 90.39, v2 60.43 on the full benchmarks), loaded with this repo as its base.

What it is

• ColPali-style multi-vector (late-interaction, MaxSim) visual document retriever, architecture ColMinistral (model_type: mistral3), built from mistralai/Ministral-3-3B-Instruct-2512: Pixtral vision tower (24 layers, patch 14, spatial merge 2) + multimodal projector + 26-layer language model + retrieval head (custom_text_proj).
• Full model in bfloat16 (model.safetensors, 7.7 GB), transformers 5.5.1 format.
• Processor (colministral3_processor_config.json): max_num_visual_tokens = 256 (the later runs train and evaluate with 1,572), query augmentation token </s>.

Use as a base

The adapters of checkpoints_diverse store, in adapter_config.json, the local path this model had during training (/home/jovyan/workplace/colbert_v1/models/colministral3). Download this repo and either recreate that path or point base_model_name_or_path to your copy:

hf download Omundu-Riad/colministral3 --local-dir /home/jovyan/workplace/colbert_v1/models/colministral3

The adapters of the latest run also contain fully trained copies of the vision tower, projector and retrieval head (modules_to_save), which replace the ones of this model when the adapter is loaded.