On this article, you’ll discover ways to use a pre-trained massive language mannequin to extract structured options from textual content and mix them with numeric columns to coach a supervised classifier.
Subjects we are going to cowl embrace:
- Making a toy dataset with combined textual content and numeric fields for classification
- Utilizing a Groq-hosted LLaMA mannequin to extract JSON options from ticket textual content with a Pydantic schema
- Coaching and evaluating a scikit-learn classifier on the engineered tabular dataset
Let’s not waste any extra time.
From Textual content to Tables: Characteristic Engineering with LLMs for Tabular Information
Picture by Editor
Introduction
Whereas massive language fashions (LLMs) are sometimes used for conversational functions in use circumstances that revolve round pure language interactions, they’ll additionally help with duties like characteristic engineering on complicated datasets. Particularly, you may leverage pre-trained LLMs from suppliers like Groq (for instance, fashions from the Llama household) to undertake knowledge transformation and preprocessing duties, together with turning unstructured knowledge like textual content into totally structured, tabular knowledge that can be utilized to gas predictive machine studying fashions.
On this article, I’ll information you thru the total technique of making use of characteristic engineering to structured textual content, turning it into tabular knowledge appropriate for a machine studying mannequin — specifically, a classifier educated on options created from textual content by utilizing an LLM.
Setup and Imports
First, we are going to make all the required imports for this sensible instance:
|
import pandas as pd import json from pydantic import BaseModel, Subject from openai import OpenAI from google.colab import userdata from sklearn.ensemble import RandomForestClassifier from sklearn.model_selection import train_test_split from sklearn.metrics import classification_report from sklearn.preprocessing import StandardScaler |
Be aware that in addition to frequent libraries for machine studying and knowledge preprocessing like scikit-learn, we import the OpenAI class — not as a result of we are going to instantly use an OpenAI mannequin, however as a result of many LLM APIs (together with Groq’s) have adopted the identical interface type and specs as OpenAI. This class due to this fact helps you work together with a wide range of suppliers and entry a variety of LLMs via a single consumer, together with Llama fashions through Groq, as we are going to see shortly.
Subsequent, we arrange a Groq consumer to allow entry to a pre-trained LLM that we are able to name through API for inference throughout execution:
|
groq_api_key = userdata.get(‘GROQ_API_KEY’) consumer = OpenAI( base_url=“https://api.groq.com/openai/v1”, api_key=groq_api_key ) |
Necessary observe: for the above code to work, you’ll want to outline an API secret key for Groq. In Google Colab, you are able to do this via the “Secrets and techniques” icon on the left-hand facet bar (this icon seems like a key). Right here, give your key the identify 'GROQ_API_KEY', then register on the Groq website to get an precise key, and paste it into the worth area.
Making a Toy Ticket Dataset
The subsequent step generates an artificial, partly random toy dataset for illustrative functions. When you have your personal textual content dataset, be at liberty to adapt the code accordingly and use your personal.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 |
import random import time
random.seed(42) classes = [“access”, “inquiry”, “software”, “billing”, “hardware”]
templates = { “entry”: [ “I’ve been locked out of my account for {days} days and need urgent help!”, “I can’t log in, it keeps saying bad password.”, “Reset my access credentials immediately.”, “My 2FA isn’t working, please help me get into my account.” ], “inquiry”: [ “When will my new credit card arrive in the mail?”, “Just checking on the status of my recent order.”, “What are your business hours on weekends?”, “Can I upgrade my current plan to the premium tier?” ], “software program”: [ “The app keeps crashing every time I try to view my transaction history.”, “Software bug: the submit button is greyed out.”, “Pages are loading incredibly slowly since the last update.”, “I’m getting a 500 Internal Server Error on the dashboard.” ], “billing”: [ “I need a refund for the extra charges on my bill.”, “Why was I billed twice this month?”, “Please update my payment method, the old card expired.”, “I didn’t authorize this $49.99 transaction.” ], “{hardware}”: [ “My hardware token is broken, I can’t log in.”, “The screen on my physical device is cracked.”, “The card reader isn’t scanning properly anymore.”, “Battery drains in 10 minutes, I need a replacement unit.” ] }
knowledge = [] for _ in vary(100): cat = random.alternative(classes) # Injecting a random variety of days into particular templates to foster selection textual content = random.alternative(templates[cat]).format(days=random.randint(1, 14))
knowledge.append({ “textual content”: textual content, “account_age_days”: random.randint(1, 2000), “prior_tickets”: random.decisions([0, 1, 2, 3, 4, 5], weights=[40, 30, 15, 10, 3, 2])[0], “label”: cat })
df = pd.DataFrame(knowledge) |
The dataset generated comprises buyer assist tickets, combining textual content descriptions with structured numeric options like account age and variety of prior tickets, in addition to a category label spanning a number of ticket classes. These labels will later be used for coaching and evaluating a classification mannequin on the finish of the method.
Extracting LLM Options
Subsequent, we outline the specified tabular options we need to extract from the textual content. The selection of options is domain-dependent and totally customizable, however you’ll use the LLM in a while to extract these fields in a constant, structured format:
|
class TicketFeatures(BaseModel): urgency_score: int = Subject(description=“Urgency of the ticket on a scale of 1 to five”) is_frustrated: int = Subject(description=“1 if the consumer expresses frustration, 0 in any other case”) |
For instance, urgency and frustration usually correlate with particular ticket sorts (e.g. entry lockouts and outages are usually extra pressing and emotionally charged than common inquiries), so these alerts may also help a downstream classifier separate classes extra successfully than uncooked textual content alone.
The subsequent operate is a key ingredient of the method, because it encapsulates the LLM integration wanted to rework a ticket’s textual content right into a JSON object that matches our schema.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 |
def extract_features(textual content: str) -> dict: # Sleep for two.5 seconds for safer use below the constraints of the 30 RPM free-tier restrict time.sleep(2.5)
schema_instructions = json.dumps(TicketFeatures.model_json_schema()) response = consumer.chat.completions.create( mannequin=“llama-3.3-70b-versatile”, messages=[ { “role”: “system”, “content”: f“You are an extraction assistant. Output ONLY valid JSON matching this schema: {schema_instructions}” }, {“role”: “user”, “content”: text} ], response_format={“kind”: “json_object”}, temperature=0.0 ) return json.masses(response.decisions[0].message.content material) |
Why does the operate return JSON objects? First, JSON is a dependable technique to ask an LLM to supply structured outputs. Second, JSON objects might be simply transformed into Pandas Sequence objects, which might then be seamlessly merged with different columns of an current DataFrame to develop into new ones. The next directions do the trick and append the brand new options, saved in engineered_features, to the remainder of the unique dataset:
|
print(“1. Extracting structured options from textual content utilizing LLM…”) engineered_features = df[“text”].apply(extract_features) features_df = pd.DataFrame(engineered_features.tolist())
X_raw = pd.concat([df.drop(columns=[“text”, “label”]), features_df], axis=1) y = df[“label”]
print(“n2. Last Engineered Tabular Dataset:”) print(X_raw) |
Here’s what the ensuing tabular knowledge seems like:
|
account_age_days prior_tickets urgency_score is_annoyed 0 564 0 5 1 1 1517 3 4 0 2 62 0 5 1 3 408 2 4 0 4 920 1 5 1 .. ... ... ... ... 95 91 2 4 1 96 884 0 4 1 97 1737 0 5 1 98 837 0 5 1 99 862 1 4 1
[100 rows x 4 columns] |
Sensible observe on price and latency: Calling an LLM as soon as per row can develop into sluggish and costly on bigger datasets. In manufacturing, you’ll normally need to (1) batch requests (course of many tickets per name, in case your supplier and immediate design permit it), (2) cache outcomes keyed by a secure identifier (or a hash of the ticket textual content) so re-runs don’t re-bill the identical examples, and (3) implement retries with backoff to deal with transient charge limits and community errors. These three practices sometimes make the pipeline sooner, cheaper, and way more dependable.
Coaching and Evaluating the Mannequin
Lastly, right here comes the machine studying pipeline, the place the up to date, totally tabular dataset is scaled, break up into coaching and take a look at subsets, and used to coach and consider a random forest classifier.
|
print(“n3. Scaling and Coaching Random Forest…”) scaler = StandardScaler() X_scaled = scaler.fit_transform(X_raw)
# Break up the info into coaching and take a look at X_train, X_test, y_train, y_test = train_test_split(X_scaled, y, test_size=0.4, random_state=42)
# Prepare a random forest classification mannequin clf = RandomForestClassifier(random_state=42) clf.match(X_train, y_train)
# Predict and Consider y_pred = clf.predict(X_test) print(“n4. Classification Report:”) print(classification_report(y_test, y_pred, zero_division=0)) |
Listed below are the classifier outcomes:
|
Classification Report: precision recall f1–rating assist
entry 0.22 0.18 0.20 11 billing 0.29 0.33 0.31 6 {hardware} 0.29 0.25 0.27 8 inquiry 1.00 1.00 1.00 8 software program 0.44 0.57 0.50 7
accuracy 0.45 40 macro avg 0.45 0.47 0.45 40 weighted avg 0.44 0.45 0.44 40 |
When you used the code for producing an artificial toy dataset, chances are you’ll get a somewhat disappointing classifier end result when it comes to accuracy, precision, recall, and so forth. That is regular: for the sake of effectivity and ease, we used a small, partly random set of 100 cases — which is usually too small (and arguably too random) to carry out effectively. The important thing right here is the method of turning uncooked textual content into significant options via the usage of a pre-trained LLM through API, which ought to work reliably.
Abstract
This text takes a mild tour via the method of turning uncooked textual content into totally tabular options for downstream machine studying modeling. The important thing trick proven alongside the best way is utilizing a pre-trained LLM to carry out inference and return structured outputs through efficient prompting.
