kairos-quality-intent

A fine-tune of FunctionGemma-270M-it that reads a video edit request written in plain words ("give it more of a film feel", "the voice is hard to hear") and names the quality intents it asks for, from a closed list of 30 ids, or none. It runs in the browser in the kaairos video editor (transformers.js, WebGPU), where it is consulted only on a request no word rule recognised. It names WHAT is wanted; the editor's own code sizes every step from the clip's measurements, and every step waits on an approval card.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. This model is a Model Derivative of Gemma; its use is subject to the same terms, including the Gemma Prohibited Use Policy. The weights were modified: LoRA adapters were trained and merged, the model was exported to ONNX fp16, and the language-model head was tied to the embedding.

How it was made

• Data. Requests and labels were written by an open-weight teacher model (Gemma 4 31B-it) from a fixed list of intent descriptions: about 3,000 requests across 12 user personas, labelled under fixed rules.
• Training. LoRA (rank 16) for 6 epochs, merged into the base in fp32. Validation, with two personas held out whole: 85 % exact.
• Export. onnxruntime-genai's builder (fp16, WebGPU), the logits sliced to the last position, the lm_head tied to the embedding through a Gemm. On 82 reference requests the ONNX file's greedy answers equal the PyTorch model's, 82/82, and its answer confidence differs by at most 0.0008.

Use

Prompt (greedy, at most 12 new tokens, stop on <end_of_turn>, tokenised without adding special tokens):

<bos><start_of_turn>developer
Name the quality intents of the request, in order, or none.<end_of_turn>
<start_of_turn>user
{request}<end_of_turn>
<start_of_turn>model

The reply is space-separated intent ids, or none. The app keeps an answer only when the product of the chosen tokens' probabilities, the end token included, is at least 0.98, a floor chosen on validation data only.

Limits

English only. It confuses neighbouring intents (a "music video" read as cinematic) and can drop the second intent of a compound request. Measured on held-out requests it was never trained on, it found the intent on 5 of 14 that had one, and wrongly added an intent on 1 of 20.