Vision Sft

180 chapters · 328 min

Quick Reference

The Consensus Recipe

Freeze the vision tower and the projector. Put LoRA on the LLM only, all-linear (the same attention + MLP target list as text-only SFT — see lora-qlora-recipes), at r=8–16, α=16–32. This is the settled default for adapting a VLM’s behavior without disturbing how it sees.

• The vision tower and projector stay frozen by default. They already encode a general visual representation; retraining them is rarely necessary and adds risk without adding capability for most tasks.
• LoRA rank runs lower than the text-only general default (r=8–16 here vs r=16–32 for text-only SFT) because the LLM-only adapter is adapting behavior, not injecting new visual knowledge.
• QLoRA is permitted only with a frozen vision tower. Quantizing the base while also unfreezing and training vision layers is unsupported and unstable — treat this as a hard pairing rule, not a tunable. If the vision tower needs to unfreeze, drop QLoRA and use bf16 LoRA instead.

When to Unfreeze

Unfreezing vision layers is a deliberate escalation, not a default decision — reach for it only when the domain shift is visual, not textual.

• Unfreeze only for visual domain shift. If the task is teaching new behavior on images the tower already understands (charts, everyday photos), the frozen-tower recipe above is sufficient. Unfreeze when the visual domain itself is unfamiliar to the tower — satellite imagery, medical scans, dense technical diagrams — and the frozen-tower recipe plateaus.
• Last-6 ViT layers is the sweet spot. Unfreezing the final six vision-transformer layers (not the whole tower) measured +1.7pt DocVQA at ~1.75x training cost over the frozen baseline. Treat six layers as the ceiling worth paying for; going further spends compute without a matched result.
• Vision LR must run 5–10x lower than the LLM LR when unfrozen. The vision tower’s pretrained representation is more fragile than the LLM’s adapter; the same LR for both risks overwriting the visual representation faster than the LLM adapter can compensate.
• High LoRA rank on the patch- embedding layer risks NaN. If patch embedding is in the unfrozen set, keep its rank low and watch early-step loss closely — one of the most fragile places to apply LoRA in a VLM.

The Two Silent Killers

Both produce a run that trains without error and without learning: the loss curve looks normal, the model doesn’t improve, and neither throws an exception — both need an explicit pre-training check, not just a clean training log.

• Image-tag/count mismatch. Every image placeholder token in the templated text must map 1:1 to a media item actually passed to the collator. A mismatch (one placeholder, zero or two images attached; or an image with no placeholder) doesn’t error in most collators — it silently misaligns image and text, and the model “trains but learns nothing.” Validate the 1:1 placeholder-to-media mapping before training starts, on every example, not just a sample. Full validation-checklist detail: references/collators-and-pitfalls.md.
• min_pixels/max_pixels resolution budget. This pair is the single most consequential hyperparameter for quality and memory in VLM SFT — more than rank, alpha, or LR. Too low silently downsamples images below what the task needs (small document text becomes unreadable even though training “succeeds”); too high blows the activation memory budget or forces too small a batch to train stably. Set it deliberately per dataset, don’t leave it at a framework default.

Unsloth Specifics

• UnslothVisionDataCollator is the collator Unsloth expects for VLM SFT — it handles the image-tag alignment and per-architecture processor contract described in references/collators-and-pitfalls.md. Don’t substitute a text-only collator for VLM data.
• finetune_vision_layers=False is required when fast_inference=True. vLLM cannot serve LoRA adapters on vision layers, so a fast- inference setup that also unfreezes vision layers fails at serve time even if training succeeds. If the recipe calls for unfreezing the last-6 ViT layers (see When to Unfreeze above), fast inference is off the table for that run — choose one or the other, not both.

Model Choice

Base VLM choice is out of scope for this skill — it lives in one place, the model catalog at finetuning-method-selection‘s references/model-catalog.md. This skill and its references describe recipes by architecture family only, never by recommending one model over another.

VLM reinforcement learning (VLM-GRPO) is reference-only in this plugin — the fragmented tooling and reward-hacking failure modes specific to VLM-RL are covered in grpo-rlvr-training, not here. This skill’s scope stops at supervised fine-tuning.

Failure Modes

The recurring mistake across every section above is treating a clean loss curve as proof the run is healthy. A normal-looking curve is consistent with both a working run and either silent killer, since the model trains on something either way — just not the aligned image-text signal when a killer is present. A flat eval score next to a normal loss curve means re-run the checklist in references/collators-and-pitfalls.md before touching any hyperparameter.

References

• references/collators-and-pitfalls.md — per- architecture collator table, dataset-format examples with image placeholders, a pre- training validation checklist, and the two- stage projector-alignment recipe as an advanced pattern.

Related skills: finetuning-method-selection routes here; lora-qlora-recipes covers the text-only LoRA fundamentals this skill specializes; grpo-rlvr-training covers VLM-RL (reference-only); dataset-curation covers image+text dataset preparation this skill doesn’t.