Prompt the Model | HPE GEN-AI

Medical Transcripts Classification:
Prompt the Model

Introduction #

Fine-tuning your model’s output involves experimentation with various prompts and settings. Understanding each component of a prompt is crucial:

Instruction: The directive for the model.
Examples: Contextual examples for the model to consider.
Input: The data the model processes.
Expected Output: The desired result from the model.

Part 3: Prompt the Model #

Link Dataset #

Open the dataset drawer, then in Load dataset, choose the hpe-ai/medical-cases-classification-tutorial dataset from the dropdown.
Ensure the dataset previews correctly, then select Load.

Configure Initial Prompt #

Experimenting with different prompts helps to determine the most effective setup. The information the model needs could be found in either the medical_specialty column or the transcription column. We’ll try both columns and see which works best.

Customize Your Prompt #

Try prompting the model first by specialty, then by transcript, using the structured data below:

Instruction

You are an expert in medicine. Identify the most relevant medical specialty for the provided transcript, responding with a single specialty from the list below:

- Cardiovascular / Pulmonary
- Orthopedic
- Neurology
- Gastroenterology
- Obstetrics / Gynecology
- Hematology - Oncology
- ENT - Otolaryngology
- Urology
- Psychiatry / Psychology
- Ophthalmology
- Pediatrics - Neonatal

Examples

Here is an example Transcription and Class pairing:
Transcript: {{transcription}}
Class: {{medical_specialty}}

Input

Transcription: {{transcription}}
Class:

Expected Output

{{medical_specialty}}

Instruction

You are an expert in medicine. Identify the most relevant medical specialty for the provided transcript.
{{transcription}}

Only respond with a specialty from the list below:

- Cardiovascular / Pulmonary
- Orthopedic
- Neurology
- Gastroenterology
- Obstetrics / Gynecology
- Hematology - Oncology
- ENT - Otolaryngology
- Urology
- Psychiatry / Psychology
- Ophthalmology
- Pediatrics - Neonatal

Examples

Here is an example Transcription and Class pairing:
Transcription: {{transcription}} 
Class: {{medical_specialty}}

Input

Transcription: {{transcription}}
Class:

Expected Output

{{medical_specialty}}

To preview the results, select Preview.

Generate Initial Output #

Select Generate to view the model’s initial responses.

Adjust Strategy #

If the output doesn’t meet expectations, consider switching models for improved accuracy.

Open the Select model drawer and then choose Llama-2-7b-chat-hf from the list. This is a larger model with better generation capabilities.
Select Load and then Generate again to see improved results.

Optimize Model Properties #

Optimize your setup by adjusting the prompt and generation properties:

Navigate back to the model properties.
Select a resource pool to view model generation properties.
Adjust the Temperature to 0.
Reduce the Max New Tokens to 10.
Select Load.
Select Generate again.

Save the Snapshot #

After achieving satisfactory results, save this snapshot.

Select Save Snapshot.
Provide a snapshot name. Make it meaningful so you can distinguish it from other configurations you’ve tried.
Select Save.

You can load this snapshot again at any time by navigating to your project’s Snapshots tab and selecting it. Then, you can repeat this process to produce a few similar snapshots to compare the outputs and find the best one.

Recap #

You have linked the dataset and tailored the prompts to classify medical transcripts.
You have iterated on the model and its settings to enhance output quality.
You have saved your model configuration as a snapshot for future use or further refinement.