Automatically create small task-specific models
Specific AI is an automatic distillation platform for SLMs. It was built to enable enterprises to instantly and easily create any amount of task specific SLMs at scale, while ensuring models demonstrate the best quality and reliability
Replacing weeks of research with a platform that delivers in days and is simple enough for domain experts to use
Before Specific AI: weeks or months
Data scientists write the prompts, process data and curate it, choose models, fine tune them and evaluate and review results. The steps require many iterations, and a lot of manual work.
After Specific AI: days
Domain experts use Specific AI to create and distill the model for deployment.
Prompt → Distill → Deploy
STEP 1
Data collection
It begins with a prompt describing the task along with any relevant data, labeled or unlabeled. Specific AI will automatically create a raw dataset out of it, leveraging soft labeling, data synthesis and manipulating public datasets.
STEP 2
Data scoring
A score is calculated for every sample in the raw dataset to determine its contribution to the distillation process. Scoring depends on the specific task (derived from the prompt), how each sample correlates with it, the distribution and semantic position of each sample relative to others as well as data characteristics (length, language, etc.).
STEP 3
Model selection
The platform identifies the open-source model that will yield the best results. The model’s architecture is chosen based on the production requirements and the task type (e.g., a model from the BERT family for intent classification, and a generative model such as LLaMa for summarization).
STEP 4
Fine tuning
The platform deploys SOTA techniques to automatically fine-tune the selected model.
STEP 5
(Optional)
Minimal outlier feedback
You can improve data scoring by requesting domain expert feedback on some of the samples. The system suggests 15 key samples where feedback will provide the highest marginal gain for score recalculation. Our data shows that within iterations the distilled model can outperform the LLM.
STEP 6
Evaluation
The platform provides a comparison between the LLM and the distilled SLM. Benchmark metrics are chosen according to the type of the task (e.g., precision, recall and F1 for classification, and LLM as a Judge and BLEU for summarizations), and are configurable according to the customer needs.
STEP 7
Model serving
Distilled models can be deployed via Specific AI or downloaded for your use wherever suits you best.
STEP 8
Get notified when it’s time for retraining
Specific AI will alert you when it’s time to retrain the model.