Lab

Projector Tuning (Multimodal)

Projector tuning is the third training route. It trains the projection layer that connects a frozen modality encoder to the language model, so workspace models gain grounded image or audio understanding.

Where it fits

The Lab has three distinct training routes. They solve different problems and are never interchangeable:

RouteWhat trainsUse it for
PEFT Tuning (LoRA)LoRA adapter weights (rank/alpha, optionally GRPO)Task and behavior specialization from examples
Doc-to-LoRA (D2L)A hypernetwork that emits adapters from documentsContext internalization: serve context from weights
Projector tuning (this page)The encoder-to-LLM projection layer, optionally plus a LoRAMultimodal grounding: image or audio inputs

How it works

A frozen pretrained encoder turns the modality input into embeddings. The projector maps those embeddings into the language model's token space. Training aligns that mapping on paired examples from a workspace data asset. The encoder itself never trains.

ParameterTypeDescription
modalityimage | audioInput modality. Image uses a vision tower, audio uses a speech encoder.
encoderstringFrozen encoder tap. Defaults per modality: SigLIP for image, Whisper for audio.
stagesstage1 | stage1+2stage1 trains the projector only. stage1+2 also jointly tunes a LoRA adapter on the language model.
lora_rank / lora_alphaintegerLoRA hyperparameters for stage 2. Only used with stage1+2.

Each run is a durable record with status, encoder, stage progress, and HybrIE job linkage. Runs execute on the HybrIE engine's multimodal training route and report through the same job stream as other Lab training.

Projector tuning requires hybrie-server 0.1.74 or newer. Older engines reject the multimodal training route and the run fails cleanly at submission.

CLI

bash
# Create an image projector run (SigLIP encoder, projector only)
stimulir lab projectors create -m image --stages stage1 --dataset <data-asset-id> -n "vision alignment v1"

# Create an audio run that also jointly tunes a LoRA (stage 2)
stimulir lab projectors create -m audio --stages stage1+2 --lora-rank 16 --lora-alpha 32 --dataset <data-asset-id>

# Inspect runs
stimulir lab projectors list --modality image --status running
stimulir lab projectors get <run-id>

# Delete a run record (in-flight HybrIE jobs run to completion)
stimulir lab projectors delete <run-id> --yes

Console

The same runs live in the Console under Lab, in the Projectors tab. The view has modality and stage switchers, run detail with HybrIE job linkage, and run deletion.