The model classifies user messages sent to a scheduling assistant. It returns one intent label that a downstream scheduling agent can use to choose the next workflow.

Example:

text

User message:
Can you cancel my appointment tomorrow?

Model output:
CANCEL_APPOINTMENTS

This type of model is useful as a routing layer in an agentic application. Instead of asking a general LLM to complete the entire task in one step, the application can first classify intent, then call the right scheduling tool or API.

Supported Intent Labels

text

LIST_APPOINTMENTS
CANCEL_APPOINTMENTS
RESCHEDULE_APPOINTMENTS
UNSUPPORTED_REQUEST
BLOCK_SLOTS
UNBLOCK_SLOTS
UPDATE_SPECIAL_SLOTS
CONVERSATIONAL_GREETING
CLOSING_CONVERSATION

How This Model Was Fine-Tuned

The fine-tuning notebook follows this process:

Load a labeled appointment-intent dataset with user messages and target intent labels.
Clean USER_INPUT and TARGET_INTENT values by stripping extra whitespace.
Create an 80/20 stratified train/test split.
Evaluate the base meta-llama/Llama-3.2-1B-Instruct model to establish baseline accuracy and F1.
Convert each training row into a chat-formatted supervised fine-tuning example:

text

system: intent-classification instructions
user: the scheduling message
assistant: the correct intent label

Fine-tune using LoRA supervised fine-tuning.
Merge the LoRA adapter into the base model.
Evaluate the fine-tuned model on the held-out test split.
Export a GGUF build for local Ollama deployment.

Model Details

Base model: meta-llama/Llama-3.2-1B-Instruct
Fine-tuning method: supervised fine-tuning with LoRA
Export type: merged model weights
GGUF/Ollama repo: vidhyakshayakannan/appointment-intent-llama32-1b-sft-GGUF

Dataset Summary

The dataset contains 301 labeled examples with:

TEST_ID
USER_INPUT
TARGET_INTENT

The fine-tuning run used an 80/20 stratified split: 240 training examples and 61 held-out test examples.

Class distribution:

Intent	Examples
`BLOCK_SLOTS`	61
`RESCHEDULE_APPOINTMENTS`	51
`LIST_APPOINTMENTS`	46
`UNBLOCK_SLOTS`	39
`UPDATE_SPECIAL_SLOTS`	35
`CANCEL_APPOINTMENTS`	32
`UNSUPPORTED_REQUEST`	25
`CONVERSATIONAL_GREETING`	6
`CLOSING_CONVERSATION`	6

Evaluation Summary

Held-out test results from the fine-tuning notebook:

Model	Test Accuracy	Macro F1	Weighted F1
Base Llama 3.2 1B Instruct	39.3%	0.1753	0.3134
Fine-tuned model	85.2%	0.8784	0.8549

The fine-tuned model substantially improves over the base model for this intent-classification task. A production system should validate the classifier with a larger dataset that reflects real scheduling traffic.

Inference Contract

The model expects a chat-style prompt with a system instruction that asks it to return exactly one intent label.

System instruction:

text

You are an intent classifier for a doctor appointment booking system.
Classify the user message into exactly one of these following intents:
LIST_APPOINTMENTS, CANCEL_APPOINTMENTS, RESCHEDULE_APPOINTMENTS, UNSUPPORTED_REQUEST, BLOCK_SLOTS, UNBLOCK_SLOTS, UPDATE_SPECIAL_SLOTS, CONVERSATIONAL_GREETING, CLOSING_CONVERSATION
Reply with ONLY the EXACT intent label. Nothing else.

User message:

I need to move my appointment from Monday to Friday.

Expected output:

RESCHEDULE_APPOINTMENTS

Run Locally with Ollama

For local Ollama usage, use the GGUF repo:

bash

ollama pull hf.co/vidhyakshayakannan/appointment-intent-llama32-1b-sft-GGUF:Q4_K_M
curl -L https://huggingface.co/vidhyakshayakannan/appointment-intent-llama32-1b-sft-GGUF/resolve/main/Modelfile -o Modelfile
ollama create appointment-intent-sft -f Modelfile
ollama run appointment-intent-sft "I need to move my appointment from Monday to Friday."

Expected output:

RESCHEDULE_APPOINTMENTS

Building a Scheduling Agent

Use this model as the intent-routing layer:

Classify the user message with the local Ollama model.
Map the label to an application route.
Collect missing scheduling fields.
Call the relevant calendar or appointment-management tool.
Ask for confirmation before changing appointments or availability.
Use fallback handling for UNSUPPORTED_REQUEST or ambiguous messages.

Limitations

This is a compact model trained on a small labeled dataset for a narrow scheduling-intent task. It is not a medical assistant and should not provide medical advice. Production deployment should include a larger dataset, stronger evaluation, monitoring, and fallback handling for ambiguous or unsupported requests.