easysellprod

🛍️ EasySell AI — SLM Tool Calling Model

A custom 101M parameter GPT model trained to output structured JSON tool calls for a WhatsApp e-commerce shopping assistant.

Architecture

• Type: Custom GPT (Decoder-only Transformer)
• Parameters: ~101M
• Layers: 14 | Heads: 12 | Embedding dim: 768
• Context length: 1024 tokens
• Vocab size: 8000 (BPE)
• Trained on: Kaggle P100, 3 epochs, loss ~0.016
• Attention: scaled_dot_product_attention (vLLM compatible)

Quick Start

pip install transformers tokenizers torch

import torch
from transformers import AutoModelForCausalLM, AutoConfig
from tokenizers import Tokenizer
from huggingface_hub import hf_hub_download

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "TanishqV03/easysellhfvllm",
    trust_remote_code=True,
)
model.eval()

# Load tokenizer
tok_path  = hf_hub_download("TanishqV03/easysellhfvllm", "tokenizer.json")
tokenizer = Tokenizer.from_file(tok_path)
EOS_ID    = tokenizer.token_to_id("[EOS]")

# Run inference
prompt = "i want bag pack"
ids    = tokenizer.encode("[BOS] " + prompt).ids
x      = torch.tensor([ids])
gen    = model.generate_tool_call(x, eos_token_id=EOS_ID)
print(tokenizer.decode(gen))
# → {"tool": "add_to_cart_tool", "parameters": {"product": "bag pack", "quantity": 1}}

Supported Tools (12)

Hinglish Support

• "mujhe shawl chahiye" → add_to_cart_tool
• "cart dikhao" → view_cart_tool
• "mobile ki link bhejo" → search_catalog_by_text_tool

vLLM Usage

from vllm import LLM
llm = LLM(model="TanishqV03/easysellhfvllm", trust_remote_code=True)

Built by Tanishq Vemaraju — EasySell AI WhatsApp e-commerce assistant.