This repository contains the qlora-v2 LoRA adapter trained on top of Qwen/Qwen2.5-7B-Instruct for a Turkish clinic appointment assistant with tool-calling behavior.

The target behavior is:

ask clarifying questions when date or time is ambiguous
emit tool calls when booking, cancelling, rescheduling, or fetching clinic information
preserve formal Turkish tone

Training summary

Base model: Qwen/Qwen2.5-7B-Instruct
Fine-tuning method: QLoRA / PEFT
Adapter type: LoRA
Runtime: 4-bit base model + adapter

Eval snapshot

Held-out eval results for the qlora-v2 setup:

Format validity: 100.0%
Tool selection: 97.0%
Argument accuracy: 91.0%
Composite tool-call accuracy: 91.0%
Tone accuracy: 100.0%

Usage

This repository contains an adapter, not a fully merged model. Load it on top of the base model:

python

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

model_name = "Qwen/Qwen2.5-7B-Instruct"
adapter_name = "Hikmet58/tr-appointment-qlora-v2-adapter"

quant = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=quant,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_name)
model.eval()

Dataset

Training and evaluation files are published separately at:

Hikmet58/tr-appointment-qlora-dataset