On this article, you’ll discover ways to use a pre-trained giant language mannequin to extract structured options from textual content and mix them with numeric columns to coach a supervised classifier.
Matters we are going to cowl embody:
- Making a toy dataset with blended 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: Function Engineering with LLMs for Tabular Information
Picture by Editor
Introduction
Whereas giant language fashions (LLMs) are usually used for conversational functions in use circumstances that revolve round pure language interactions, they’ll additionally help with duties like function engineering on complicated datasets. Particularly, you may leverage pre-trained LLMs from suppliers like Groq (for instance, fashions from the Llama household) to undertake information transformation and preprocessing duties, together with turning unstructured information like textual content into totally structured, tabular information that can be utilized to gas predictive machine studying fashions.
On this article, I’ll information you thru the total means of making use of function engineering to structured textual content, turning it into tabular information appropriate for a machine studying mannequin — particularly, a classifier skilled on options created from textual content by utilizing an LLM.
Setup and Imports
First, we are going to make all the mandatory imports for this sensible instance:
|
import pandas as pd import json from pydantic import BaseModel, Area 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 |
Notice that in addition to widespread libraries for machine studying and information preprocessing like scikit-learn, we import the OpenAI class — not as a result of we are going to immediately 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 subsequently helps you work together with a wide range of suppliers and entry a variety of LLMs by way of 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 will 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 ) |
Essential notice: for the above code to work, you should outline an API secret key for Groq. In Google Colab, you are able to do this by way of the “Secrets and techniques” icon on the left-hand aspect bar (this icon appears like a key). Right here, give your key the identify 'GROQ_API_KEY', then register on the Groq web site to get an precise key, and paste it into the worth discipline.
Making a Toy Ticket Dataset
The following step generates an artificial, partly random toy dataset for illustrative functions. When you have your personal textual content dataset, be happy 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.” ] }
information = [] for _ in vary(100): cat = random.selection(classes) # Injecting a random variety of days into particular templates to foster selection textual content = random.selection(templates[cat]).format(days=random.randint(1, 14))
information.append({ “textual content”: textual content, “account_age_days”: random.randint(1, 2000), “prior_tickets”: random.selections([0, 1, 2, 3, 4, 5], weights=[40, 30, 15, 10, 3, 2])[0], “label”: cat })
df = pd.DataFrame(information) |
The dataset generated comprises buyer help 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 wish to extract from the textual content. The selection of options is domain-dependent and totally customizable, however you’ll use the LLM afterward to extract these fields in a constant, structured format:
|
class TicketFeatures(BaseModel): urgency_score: int = Area(description=“Urgency of the ticket on a scale of 1 to five”) is_frustrated: int = Area(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 basic inquiries), so these alerts might help a downstream classifier separate classes extra successfully than uncooked textual content alone.
The following perform is a key aspect of the method, because it encapsulates the LLM integration wanted to remodel 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 underneath 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.hundreds(response.selections[0].message.content material) |
Why does the perform return JSON objects? First, JSON is a dependable solution to ask an LLM to supply structured outputs. Second, JSON objects may be simply transformed into Pandas Sequence objects, which may then be seamlessly merged with different columns of an present DataFrame to turn out to be 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. Ultimate Engineered Tabular Dataset:”) print(X_raw) |
Here’s what the ensuing tabular information appears like:
|
account_age_days prior_tickets urgency_score is_pissed off 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 notice on price and latency: Calling an LLM as soon as per row can turn out to be gradual and costly on bigger datasets. In manufacturing, you’ll often wish to (1) batch requests (course of many tickets per name, in case your supplier and immediate design enable 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 fee limits and community errors. These three practices usually make the pipeline quicker, 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, cut up into coaching and check 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)
# Cut up the information into coaching and check 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 help
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 |
If you happen to used the code for producing an artificial toy dataset, chances are you’ll get a reasonably disappointing classifier consequence 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 often 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 by way of the usage of a pre-trained LLM through API, which ought to work reliably.
Abstract
This text takes a delicate tour by way of 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.
