vision-sft – skill Thiết kế & UI cho agent AI

Vision-Language SFT

This skill assumes finetuning-method-selection

already routed here: the data shape is

image+text demonstrations, not preference pairs

or a verifiable reward signal, and the base is a

vision-language model rather than a text-only

one. lora-qlora-recipes covers the text-only

LoRA/QLoRA recipe this skill specializes for the

vision tower and projector; read that skill first

if the LoRA fundamentals (rank, alpha, target

modules) aren't already familiar.

Input: an image+text dataset and a VLM base

model already picked from the model catalog.

Output format: a validated adapter config —

which components are frozen, LoRA target modules,

and a min_pixels/max_pixels budget — that

llm-finetuning-training-engineer consumes

directly when it generates a runnable script.

Quick Reference

| Situation | Default |

|---|---|

| Adapting behavior on familiar images | Frozen tower+projector, LoRA r=8–16, α=16–32 |

| Visual domain shift | Unfreeze last-6 ViT layers, vision LR 5–10x lower |

| Doesn't fit in bf16 at target rank | QLoRA — frozen vision tower only |

| fast_inference=True | finetune_vision_layers=False |

| Loss normal, eval not improving | Check the Two Silent Killers below first |

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.

# freeze tower + projector; LoRA on LLM only
for name, param in model.named_parameters():
    if "vision_tower" in name or "projector" in name:
        param.requires_grad = False

target_modules = [
    "q_proj", "k_proj", "v_proj", "o_proj",
    "gate_proj", "up_proj", "down_proj",
]  # LLM-only, all-linear — r=8-16, alpha=16-32

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.