Model Card for Waiter-0.8B-LoRA

This model is a QLoRA fine-tuned version of Qwen/Qwen3.5-0.8B for multilingual Adisyonist POS tool routing.

Fine-tuned by Ömer Faruk ŞAHAN — Adisyonist AI

Full merged model:

adisyonist/waiter-0.8B

Quick Start

from transformers import AutoModelForImageTextToText
from peft import PeftModel

base_model = AutoModelForImageTextToText.from_pretrained(
    "Qwen/Qwen3.5-0.8B",
    device_map="auto",
    torch_dtype="auto"
)

model = PeftModel.from_pretrained(
    base_model,
    "adisyonist/waiter-0.8B-lora"
)
'''

## Output Format

The model generates a single JSON object:

```json
{
  "agents": ["prepareAddItems", "executeTableAction"],
  "answer": "..."
}

For requests that do not require a tool:

{
  "agents": [],
  "answer": "..."
}

Training Procedure

SFT

The model was fine-tuned using QLoRA with TRL SFTTrainer.

• Base model: Qwen/Qwen3.5-0.8B
• Method: QLoRA
• Quantization: 4-bit NF4
• Double quantization: Enabled
• Compute dtype: BF16
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Corpus size: 532,856 examples
• Training split: 426,285 examples
• Validation split: 53,285 examples
• Held-out test split: 53,286 examples
• Languages: 43

Evaluation

Held-out test results:

• Raw agent exact accuracy: 98.81%
• Adjusted agent exact accuracy: 99.40%
• JSON validity: 99.41%
• Answer exact-match: 92.57%

The adjusted routing score accounts for outputs where generation was truncated by the configured token limit while the generated agent chain itself was correct.

Frameworks

• Transformers
• PEFT
• TRL
• PyTorch
• bitsandbytes

License

The adapter is based on Qwen/Qwen3.5-0.8B and is released under the Apache License 2.0.