---

name: training description: "Guides MiniMind-V Pretrain and SFT command planning, DDP launch, freezing, checkpoint/resume, precision, logging, prerequisites, and output naming." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MiniMind-V Training

Use this sub-skill when the task is to plan, configure, resume, or safely construct commands for MiniMind-V VLM Pretrain or SFT training.

Do not run full training as a default verification step. MiniMind-V training is GPU-, data-, and weight-gated; build commands, check prerequisites, and ask before launching any expensive training job.

Route here

• Choosing between optional Pretrain and SFT-first training.
• Constructing single-GPU or DDP torchrun commands from a MiniMind-V checkout.
• Setting freeze_llm, from_weight, from_resume, precision, logging, and checkpoint options.
• Understanding output weight names, resume state, and GPU-count-change resume behavior.
• Troubleshooting missing weights/data, CUDA/OOM, tokenizer/model resources, DDP ranks, checkpoint mismatch, logging, or torch.compile.

Route elsewhere

• Parquet schema or resource downloads: data-and-resources.
• Model internals or SigLIP2 architecture: model-architecture-and-api.
• Inference, evaluation, WebUI, or serving: inference-and-serving.
• Post-training conversion to Transformers: model-export-and-format-conversion.

Operating workflow

1. Identify the desired stage: sft for the recommended main path or pretrain for optional projector alignment before SFT.
2. Confirm prerequisites without downloading them: parquet data, base/previous-stage weights, tokenizer files, and SigLIP2.
3. Use build_training_command.py for safe planning; it prints commands and optional file-check results but never launches training.
4. Read training workflows, CLI reference, checkpoints and resume, and troubleshooting.

Key defaults

• Pretrain: save_weight=pretrain_vlm, batch_size=16, learning_rate=4e-4, max_seq_len=450, data_path=../dataset/pretrain_i2t.parquet, from_weight=llm, freeze_llm=2.
• SFT: save_weight=sft_vlm, batch_size=4, learning_rate=5e-6, max_seq_len=768, data_path=../dataset/sft_i2t.parquet, from_weight=pretrain_vlm, freeze_llm=1.
• To skip Pretrain, use SFT with --from_weight llm because the SFT parquet already includes the Pretrain caption subset.