README.md

metadata

pretty_name: ColMinistral3 training checkpoints
tags:
  - colpali
  - visual-document-retrieval
  - late-interaction
  - multi-vector
  - peft
  - dora

ColMinistral3 training checkpoints (checkpoints_diverse)

Training checkpoints of ColMinistral3: a ColPali-style multi-vector (late-interaction) visual document retriever built on mistralai/Ministral-3-3B-Instruct-2512. A page image and a text query are both encoded by the vision-language model into one 320-d vector per token, and a query-page score is the ColBERT MaxSim (sum over query tokens of the best-matching page token).

This repository is a storage of experiment checkpoints (it uses the dataset repo type only as a file store; it is not a dataset). Every checkpoint is a PEFT adapter (LoRA / DoRA / DeLoRA) for the base model, plus — depending on the run — the fully trained vision tower, projector and retrieval head.

Which checkpoint to use: colministral_siglip_vision_weighted16/best_vidore_v1_step5200 — the best model measured on the full ViDoRe v1 and v2 benchmarks (see below).

Folder layout

<run name>/<checkpoint>/. Files you will find in a checkpoint folder:

File	Content
adapter_model.safetensors	trained weights: PEFT adapter of the language model, plus every module listed in modules_to_save of adapter_config.json (fully trained copies)
adapter_config.json	PEFT config: adapter type, rank r, lora_alpha, target_modules, modules_to_save, base model path
optimizer.pt, scheduler.pt, rng_state*.pth	optimizer / LR-scheduler / RNG state: present only in raw checkpoint-N folders, needed to resume training
trainer_state.json	Hugging Face Trainer state: global step, full per-step training log (log_history)
training_args.bin	Hugging Face TrainingArguments (pickle)
processor_config.json, colministral3_processor_config.json, tokenizer*.json, chat_template.jinja	processor: image budget (max_num_visual_tokens), tokenizer, chat template
eval_vidore.json	(best-adapter folders) the in-training ViDoRe lite scores that selected this adapter
vidore_results/	mteb result JSONs of the full ViDoRe benchmarks for this checkpoint

Latest run: colministral_siglip_vision_weighted16

Continued training of ColMinistral3 on 16 document domains with manual sampling weights, one domain per batch.

• Initialisation: end of the previous phase (siglip_vision_phase2/checkpoint-5400, itself continued from siglip_vision_500steps/checkpoint-500; not in this repo).
• Data: Omundu-Riad/vqa_train_clean — 737,626 query-page pairs in 16 domains (ArxivQA, DocVQA, InfoVQA, MLPapers, TatDQA, WebPDF, PlotQA, ChartQA, SlideVQA, Manuals, Textbooks, VDR-en/fr/de/es/it), deduplicated and filtered against the ViDoRe benchmarks (no benchmark page or query is used for training). Domain weights: TatDQA 3.0, SlideVQA 3.0, WebPDF 2.0, ChartQA 2.0, VDR-en 1.8, DocVQA / InfoVQA / Textbooks 1.5, ArxivQA and VDR-fr/de/es/it 1.0, MLPapers 0.8, Manuals 0.75, PlotQA 0.7.
• Trained parameters: DoRA (r = 32, alpha = 32, dropout 0.1) on all attention and MLP projections of the 26 language-model layers; vision tower, multimodal projector, retrieval head (custom_text_proj, 320-d) and SigLIP logit scale/bias fully trained (modules_to_save).
• Loss: pairwise SigLIP (sigmoid) loss on MaxSim scores with in-batch negatives; two pairs that share the same page are both positives.
• Schedule: 1 epoch = 6,798 steps, global batch 128 (2 × A100 40 GB, mini-batches of 8), learning rates 2e-5 (DoRA + head) / 5e-6 (vision tower + projector) / 1e-3 (SigLIP logit), 100 warm-up steps, gradient clipping 1.0, bf16 autocast. No query augmentation tokens. Images: up to 1,572 visual tokens per page.
• Selection: a ViDoRe v1 + v2 lite evaluation (subsampled corpora) every 400 steps; the best adapter per benchmark is kept (best_vidore_v1_*, best_vidore_v2_*). The lite history is eval_vidore.jsonl in the run folder, the full training log is training.log.

Checkpoints and full ViDoRe scores

Official mteb main scores (mteb 2.3.2): nDCG@5 × 100 for ViDoRe v1 (10 tasks) and v2 (4 tasks, averaged over query languages), nDCG@10 × 100 for ViDoRe v3 (the 8 public tasks).

Folder	Step	ViDoRe v1	ViDoRe v2	ViDoRe v3	Can resume training
(start of the run, not in this repo)	0	88.42	56.77	—	—
best_vidore_v2_step3600	3600	89.77	60.37	—	no (adapter only)
best_vidore_v1_step5200	5200	90.39	60.43	58.49	no (adapter only)
checkpoint-6798 (end of the run)	6798	90.24	57.83	—	yes (optimizer, scheduler, RNG)

• best_vidore_v1_step5200 is the best checkpoint on both full v1 and v2. After step 5200, v1 stays flat and v2 drops (−2.6 at the end of the run), mostly on long financial / ESG reports (the right page is still in the top 10, but less often ranked first).
• best_vidore_v2_step3600 was selected by the lite v2 score; on the full benchmark it ties step 5200 on v2.
• The optimizer state of steps 3600 and 5200 was not kept (only the 2 latest raw checkpoints are kept during training), so these two folders are inference-ready adapters, not resumable checkpoints.

Query augmentation tokens (evaluation-only probe)

The model was trained without query augmentation tokens. Appending N </s> tokens to every query at evaluation time only, on best_vidore_v1_step5200 (result JSONs in vidore_results_query_aug<N>/):

</s> tokens appended to the query	ViDoRe v1	ViDoRe v2
0 (as trained)	90.39	60.43
5	90.56	61.21
10	90.75	61.36
20	90.54	63.38

Loading

The adapters target the ColMinistral3 architecture of the training code (a colpali_engine fork with ColMinistral and ColMinistral3Processor). adapter_config.json points to the local path of the base model used in training (the legacy ColMinistral3 model, on the Hub as Omundu-Riad/colministral3); download it and recreate that path or set base_model_name_or_path to your copy. Evaluation reproduces training exactly: frozen base in bf16, unmerged DoRA adapter, trained modules in fp32, forward under bf16 autocast, every non-padding token is a vector, MaxSim summed over query tokens. Raw checkpoint-N folders of this run have no processor files; checkpoint-6798 here includes them (copied from the run's final output, whose adapter is identical to checkpoint-6798). Image budget: max_num_visual_tokens=1572.

Older runs

Earlier experiments, all adapters of mistralai/Ministral-3-3B-Instruct-2512. The description below is read from each folder's own adapter_config.json and trainer_state.json:

• Adapter: PEFT type, rank r / lora_alpha. LLM = attention and MLP projections of the language model + retrieval head (custom_text_proj); LLM + vision q,v also adapts the q/v projections of the vision tower.
• Step: global step / planned total steps of the run; batch = training batch size recorded by the Trainer.
• Train loss: mean of the last 50 logged training losses. Eval loss: the run's own in-training validation loss at that step. Losses depend on the run's loss function and batch size: compare them within a run, not across runs.
• Resume: whether the folder has everything the Trainer needs to resume (optimizer.pt, scheduler.pt, trainer_state.json).

Folder	Adapter	Target	Step	Epoch	Batch	Peak LR	Train loss	Eval loss	Resume	Size
colministral_DORA/checkpoint-1500	DoRA r=64/α=64	LLM	1500 / 3694	0.406	32	5.0e-05	0.1618	0.1946	yes	1.2 GB
colministral_DORA/checkpoint-2000	DoRA r=64/α=64	LLM	2000 / 3694	0.541	32	5.0e-05	0.1519	0.2003	yes	1.2 GB
colministral_DORA_128b_lr35/checkpoint-1100	DoRA r=64/α=64	LLM	1100 / 4620	1.19	128	3.0e-05	0.2574	0.3026	yes	1.2 GB
colministral_DORA_128b_lr35/checkpoint-600	DoRA r=64/α=64	LLM	600 / 4620	0.649	128	3.0e-05	0.3043	0.2914	yes	1.2 GB
colministral_DORA_128b_lr55/checkpoint-1000	DoRA r=64/α=64	LLM	1000 / 4620	1.082	128	5.0e-05	0.2639	0.2892	yes	1.2 GB
colministral_DORA_128b_lr55/checkpoint-1100	DoRA r=64/α=64	LLM	1100 / 4620	1.19	128	5.0e-05	0.2553	0.2931	yes	1.2 GB
colministral_DORA_128b_lr75/checkpoint-1000	DoRA r=64/α=64	LLM	1000 / 4620	1.082	128	7.0e-05	0.2669	0.2838	yes	1.2 GB
colministral_DORA_128b_lr75/checkpoint-1100	DoRA r=64/α=64	LLM	1100 / 4620	1.19	128	7.0e-05	0.2598	0.2801	yes	1.2 GB
colministral_DORA_128b_lr95/checkpoint-1000	DoRA r=64/α=64	LLM	1000 / 4620	1.082	128	9.0e-05	0.2748	0.2936	yes	1.2 GB
colministral_DORA_128b_lr95/checkpoint-1100	DoRA r=64/α=64	LLM	1100 / 4620	1.19	128	9.0e-05	0.2661	0.3045	yes	1.2 GB
colministral_DORA_32b_lr56_hardDs_vision_r32/checkpoint-3800	DoRA r=32/α=32	LLM + vision q,v	3800 / 9770	1.945	32	5.0e-06	0.2612	0.1711	yes	0.64 GB
colministral_DORA_32b_lr56_hardDs_vision_r32/checkpoint-4000	DoRA r=32/α=32	LLM + vision q,v	4000 / 9770	2.047	32	5.0e-06	0.2584	0.1915	yes	0.64 GB
colministral_DORA_32b_lr75_hardDs/checkpoint-2300	DoRA r=64/α=64	LLM	2300 / 9770	1.177	32	7.0e-05	0.2881	0.2069	yes	1.2 GB
colministral_DORA_32b_lr75_hardDs/checkpoint-3000	DoRA r=64/α=64	LLM	3000 / 9770	1.535	32	7.0e-05	0.2724	0.2101	yes	1.2 GB
colministral_DORA_32b_lr75_hardDs_vision/checkpoint-3800	DoRA r=32/α=32	LLM + vision q,v	3800 / 9770	1.945	32	7.0e-05	0.2518	0.1747	yes	0.64 GB
colministral_DORA_32b_lr75_hardDs_vision/checkpoint-4900	DoRA r=32/α=32	LLM + vision q,v	4900 / 9770	2.508	32	7.0e-05	0.2082	0.2217	yes	0.64 GB
colministral_DORA_32b_lr75_hardDs_vision_r16/checkpoint-3100	DoRA r=16/α=16	LLM + vision q,v	3100 / 9770	1.586	32	7.0e-05	0.2733	0.1701	yes	0.33 GB
colministral_DORA_32b_lr75_hardDs_vision_r16/checkpoint-5500	DoRA r=16/α=16	LLM + vision q,v	5500 / 9770	2.815	32	7.0e-05	0.212	0.1833	yes	0.33 GB
colministral_DORA_complete_wrapper_fix_64b/checkpoint-1500	DoRA r=64/α=64	LLM	1500 / 5541	0.812	64	2.0e-05	0.1884	0.2280	yes	1.2 GB
colministral_DORA_complete_wrapper_fix_64b/checkpoint-2000	DoRA r=64/α=64	LLM	2000 / 5541	1.083	64	2.0e-05	0.1865	0.2033	yes	1.2 GB
colministral_DORA_complete_wrapper_fix_64b_lr5/checkpoint-1500	DoRA r=64/α=64	LLM	1500 / 5541	0.812	64	5.0e-05	0.192	0.2251	yes	1.2 GB
colministral_DORA_complete_wrapper_fix_64b_lr5/checkpoint-2000	DoRA r=64/α=64	LLM	2000 / 5541	1.083	64	5.0e-05	0.1828	0.2262	yes	1.2 GB
colministral_DORA_complete_wrapper_fix_pad/checkpoint-3500	DoRA r=64/α=64	LLM	3500 / 11082	0.947	32	2.0e-05	0.1324	0.1669	yes	1.2 GB
colministral_DORA_complete_wrapper_fix_pad/checkpoint-5500	DoRA r=64/α=64	LLM	5500 / 11082	1.489	32	2.0e-05	0.1027	0.1756	yes	1.2 GB
colministral_DORA_offical_wrraper_run2/checkpoint-1500	DoRA r=64/α=64	LLM	1500 / 11082	0.406	32	2.0e-05	0.1528	0.1632	yes	1.2 GB
colministral_DORA_offical_wrraper_run2/checkpoint-4000	DoRA r=64/α=64	LLM	4000 / 11082	1.083	32	2.0e-05	0.12	0.1712	yes	1.2 GB
colministral_DORA_offical_wrraper_run2/checkpoint-4500	DoRA r=64/α=64	LLM	n/a	—	—	—	—	—	no	0.67 GB
colministral_delora_resize	DeLoRA r=64	LLM	final adapter	—	—	—	—	—	no	0.42 GB
colministral_delora_resize/checkpoint-1000	DeLoRA r=64	LLM	1000 / 1847	0.541	64	2.0e-05	0.2324	0.2532	yes	1.19 GB
colministral_delora_resize/checkpoint-1847	DeLoRA r=64	LLM	1847 / 1847	1.0	64	2.0e-05	0.2028	—	yes	1.19 GB
colministral_instruct	DeLoRA r=64	LLM	final adapter	—	—	—	—	—	no	0.42 GB
colministral_instruct/checkpoint-1000	DeLoRA r=64	LLM	1000 / 1847	0.541	64	2.0e-05	0.2313	0.2610	yes	1.19 GB
colministral_instruct/checkpoint-1847	DeLoRA r=64	LLM	1847 / 1847	1.0	64	2.0e-05	0.2038	—	yes	1.19 GB
colministral_left_lora/checkpoint-1000	LoRA r=32/α=32	LLM	1000 / 3694	0.271	32	5.0e-04	0.6161	0.5700	yes	0.59 GB
colministral_left_lora/checkpoint-1500	LoRA r=32/α=32	LLM	1500 / 3694	0.406	32	5.0e-04	0.5926	0.6067	yes	0.59 GB
colministral_lora/checkpoint-1000	LoRA r=32/α=32	LLM	1000 / 3694	0.271	32	5.0e-04	0.8159	0.6725	yes	0.59 GB
colministral_lora/checkpoint-1500	LoRA r=32/α=32	LLM	1500 / 3694	0.406	32	5.0e-04	1.1435	1.7930	yes	0.59 GB
colministral_resize_lora	LoRA r=32/α=32	LLM	final adapter	—	—	—	—	—	no	0.22 GB

Notes on the older runs:

• Folders without a checkpoint-N suffix (colministral_delora_resize, colministral_instruct, colministral_resize_lora) are the final adapter of the run, with processor files (max_num_visual_tokens = 1572).
• colministral_DORA_offical_wrraper_run2/checkpoint-4500 is incomplete: no trainer_state.json, scheduler.pt or RNG state, and its optimizer.pt is smaller than in the other checkpoints of the run (partial upload); .swp files are editor leftovers.
• .ipynb_checkpoints/ sub-folders are notebook autosave leftovers, not checkpoints.

Provenance

Trained on 2 × A100 40 GB. None of the ViDoRe benchmark data was used for training the latest run; the benchmarks are only used for evaluation, with mteb 2.3.2.