Product Analysis — WINNER / LOSER / RISKY verdicts with red flags, market truth, and investment recommendation
Ad Copy Generation — hooks, headlines, primary text, and CTA for Meta Ads, TikTok, Google

Training

Base model: meta-llama/Meta-Llama-3.1-8B-Instruct
Method: QLoRA (4-bit NF4, LoRA rank 16, alpha 32)
Precision: bf16
Dataset: 150,000 examples (75K product analysis + 75K marketing copy)
Source: McAuley-Lab/Amazon-Reviews-2023 — 25 categories, 135,235 real products with reviews
Training loss: ~0.78 | Token accuracy: ~0.84

What each training example contains

PRODUCT: [real product title]
CATEGORY: [category]
PRICE: [price]
AVERAGE RATING: [X/5] (N reviews)
DESCRIPTION: [real product description]
FEATURES: [real product features]
CUSTOMER REVIEWS:
[★☆☆☆☆] [real review text]
[★★★★★] [real review text]
→ VERDICT: WINNER/LOSER/RISKY + full analysis

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

BASE_MODEL = "meta-llama/Meta-Llama-3.1-8B-Instruct"
LORA_MODEL = "PanosAnag/llama-3.1-ecommerce-analyzer"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

tokenizer = AutoTokenizer.from_pretrained(LORA_MODEL)
model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, quantization_config=bnb_config, device_map="auto")
model = PeftModel.from_pretrained(model, LORA_MODEL)
model.eval()

GitHub

https://github.com/PanagiotisAnag/llama-3.1-ecommerce-analyzer

Framework Versions

PEFT 0.19.1
TRL 0.24.0
Transformers 5.5.0
PyTorch 2.6.0+cu124