adisyonist/waiter-1B
adisyonist/waiter-1B is a 1 billion parameter chat agent and tool-router model developed by Adisyonist, fine-tuned from google/gemma-3-1b-it. Optimized for restaurant point-of-sale (POS) scenarios, it processes user messages and outputs a JSON object containing backend function calls and a user-facing reply. This model is specifically designed for tasks like order taking, table management, and reservation processing within the Adisyonist POS system, featuring a 32768 token context length.
Loading preview...
Overview
adisyonist/waiter-1B is a specialized 1 billion parameter language model developed by Adisyonist, built upon the google/gemma-3-1b-it base model. It functions as a chat agent and tool-router, specifically designed for restaurant point-of-sale (POS) systems. The model was fine-tuned using QLoRA on a custom dataset, adisyonist/adisyonist-chat, which includes both Turkish and English subsets.
Key Capabilities
- Structured JSON Output: The model's primary output is a JSON object containing an
agentsarray (backend functions to call) and areplystring for the user. This enables seamless integration with application logic. - POS Scenario Optimization: It is highly specialized for tasks within a restaurant POS environment, including order taking, table management, reservation handling, inventory, staff operations, and sales summaries.
- Tool Routing: Identifies and suggests the appropriate backend functions (
agents) to execute based on user requests.
Training Details
The model was fine-tuned using QLoRA (4-bit NF4, LoRA r=32, alpha=64) across all attention and MLP projections. Training utilized TRL SFTTrainer with NEFTune (alpha=5), a cosine learning rate schedule, and early stopping based on evaluation loss.
Limitations
- Domain Specificity: This model is not intended for general-purpose assistant use; its capabilities are strictly tailored to Adisyonist POS scenarios.
- JSON-only Output: It exclusively produces JSON output, and users should not expect plain text responses.