shashank ms

Posted on Sep 6

Transfer Learning in LLM Models: A Comprehensive Guide

#learnai #oxlo #ai

Most teams do not need to fine-tune a foundation model. They need to transfer its general reasoning into a specific domain. In this guide we will build a support-ticket classifier that adapts Llama 3.3 70B to a custom company taxonomy using few-shot prompt transfer. No training clusters, no weight updates, just an Oxlo.ai API key and a handful of labeled examples.

What you'll need

Python 3.10 or newer
The OpenAI SDK: pip install openai
An Oxlo.ai API key from https://portal.oxlo.ai

Step 1: Set Up the Oxlo.ai Client

I start every project with a thin wrapper around the OpenAI SDK pointed at Oxlo.ai. Because Oxlo.ai is fully OpenAI-compatible, the only difference is the base URL.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.oxlo.ai/v1",
    api_key=os.environ.get("OXLO_API_KEY", "YOUR_OXLO_API_KEY"),
)

Step 2: Curate the Domain Examples

Transfer learning needs a signal. I collected eight real support tickets and mapped them to our internal categories: billing, integration, bug, and feature_request. These examples teach the model our vocabulary.

EXAMPLES = [
    {"text": "I was charged twice for the Pro plan this month.", "label": "billing"},
    {"text": "Your webhook keeps returning 403 to our staging server.", "label": "integration"},
    {"text": "The export button crashes when I select CSV format.", "label": "bug"},
    {"text": "Can you add dark mode to the dashboard?", "label": "feature_request"},
    {"text": "My invoice shows the wrong VAT number.", "label": "billing"},
    {"text": "How do I rotate my API key inside Terraform?", "label": "integration"},
    {"text": "Notifications are sent twice for the same event.", "label": "bug"},
    {"text": "It would be great to have SSO via OIDC.", "label": "feature_request"},
]

Step 3: Define the System Prompt

The system prompt anchors the model. It restricts output to our taxonomy and prevents the model from adding conversational filler.

SYSTEM_PROMPT = """You are a support-ticket classifier for a B2B SaaS platform.
Your job is to read the user's message and return exactly one label from this list:
billing, integration, bug, feature_request.

Rules:
- Return only the label, no punctuation, no explanation.
- If a ticket matches multiple categories, pick the dominant one.
- Base your decision on the examples provided in the conversation."""

Step 4: Assemble the Few-Shot Prompt

Now I wire the examples and the new ticket into a single message list. Because Oxlo.ai charges per request rather than per token, I can pack all eight examples into the context window without worrying about input length driving up cost. That is the main reason I reach for Oxlo.ai when prototyping transfer-learning workflows.

def build_messages(ticket_text: str) -> list[dict]:
    messages = [{"role": "system", "content": SYSTEM_PROMPT}]
    for ex in EXAMPLES:
        messages.append({"role": "user", "content": ex["text"]})
        messages.append({"role": "assistant", "content": ex["label"]})
    messages.append({"role": "user", "content": ticket_text})
    return messages

Step 5: Run Inference

Finally, I send the assembled conversation to Llama 3.3 70B. I keep temperature low because classification is a deterministic task.

def classify_ticket(ticket_text: str) -> str:
    messages = build_messages(ticket_text)
    response = client.chat.completions.create(
        model="llama-3.3-70b",
        messages=messages,
        temperature=0.1,
        max_tokens=20,
    )
    return response.choices[0].message.content.strip()

Step 6: Batch Test

I run a small holdout set to verify the transfer worked. Here are three unseen tickets.

TEST_TICKETS = [
    "We need an invoice for last quarter for our finance team.",
    "The Python SDK throws a KeyError on line 42 when parsing the response.",
    "Please support SCIM user provisioning.",
]

for t in TEST_TICKETS:
    label = classify_ticket(t)
    print(f"Ticket: {t}\nLabel: {label}\n")

Run It

Putting it all together, the script looks like this. Save it as classify.py, export your key, and run python classify.py.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.oxlo.ai/v1",
    api_key=os.environ.get("OXLO_API_KEY", "YOUR_OXLO_API_KEY"),
)

EXAMPLES = [
    {"text": "I was charged twice for the Pro plan this month.", "label": "billing"},
    {"text": "Your webhook keeps returning 403 to our staging server.", "label": "integration"},
    {"text": "The export button crashes when I select CSV format.", "label": "bug"},
    {"text": "Can you add dark mode to the dashboard?", "label": "feature_request"},
    {"text": "My invoice shows the wrong VAT number.", "label": "billing"},
    {"text": "How do I rotate my API key inside Terraform?", "label": "integration"},
    {"text": "Notifications are sent twice for the same event.", "label": "bug"},
    {"text": "It would be great to have SSO via OIDC.", "label": "feature_request"},
]

SYSTEM_PROMPT = """You are a support-ticket classifier for a B2B SaaS platform.
Your job is to read the user's message and return exactly one label from this list:
billing, integration, bug, feature_request.

Rules:
- Return only the label, no punctuation, no explanation.
- If a ticket matches multiple categories, pick the dominant one.
- Base your decision on the examples provided in the conversation."""

def build_messages(ticket_text: str) -> list[dict]:
    messages = [{"role": "system", "content": SYSTEM_PROMPT}]
    for ex in EXAMPLES:
        messages.append({"role": "user", "content": ex["text"]})
        messages.append({"role": "assistant", "content": ex["label"]})
    messages.append({"role": "user", "content": ticket_text})
    return messages

def classify_ticket(ticket_text: str) -> str:
    messages = build_messages(ticket_text)
    response = client.chat.completions.create(
        model="llama-3.3-70b",
        messages=messages,
        temperature=0.1,
        max_tokens=20,
    )
    return response.choices[0].message.content.strip()

TEST_TICKETS = [
    "We need an invoice for last quarter for our finance team.",
    "The Python SDK throws a KeyError on line 42 when parsing the response.",
    "Please support SCIM user provisioning.",
]

for t in TEST_TICKETS:
    label = classify_ticket(t)
    print(f"Ticket: {t}\nLabel: {label}\n")

Expected output:

Ticket: We need an invoice for last quarter for our finance team.
Label: billing

Ticket: The Python SDK throws a KeyError on line 42 when parsing the response.
Label: bug

Ticket: Please support SCIM user provisioning.
Label: feature_request

Wrap-Up

This is the cheapest form of transfer learning: no gradients, no infrastructure, just structured context. If the volume grows, the next step is to swap the static example list for dynamic retrieval from a vector store, or to move to a longer-context model on Oxlo.ai such as Kimi K2.6 so you can include larger example banks. You could also pipe the output directly into a webhook to route tickets in real time.

DEV Community