SahayakLLM: an adapter for e-commerce support
1 min readupdated
This QLoRA adapter (a small add-on layer for a base model) built on Qwen/Qwen2.5-1.5B-Instruct is fine-tuned for two tasks: generating customer support responses and classifying requests in JSON format. The model is based on the Bitext Customer Support dataset and optimized with 4-bit quantization. This adapter is useful for developers of automated support systems and e-commerce chat bots.
Key features
- Dual tasks: support response generation and intent routing
- Trained on a dataset of 23,875 cleaned entries
- Uses QLoRA with rank r=16 and an alpha parameter of 32
- High macro-F1 scores for intents on the test split
What you can do
- Integrate it into a chat bot to automatically answer customer questions
- Use it to classify inquiries into structured JSON
- Adapt it for real company data using Qwen2.5-1.5B
Example
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.float16)
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", quantization_config=bnb, device_map={"": 0})
model = PeftModel.from_pretrained(base, "aryan22medigeri/llm-qwen2.5-1.5b-qlora")Read more
DetailsBenchmarks (C2): Intent macro-F1 0.999, BERTScore-F1 0.923, ROUGE-L 0.445
Requirements: 1.5B parameters, 4-bit quantization (NF4)
Limitations: English language only, synthetic data from fictional company KharidiKart, placeholders require tool verification










Comments