yuchenlin

"""
Decompose and memorize by program, only asking LLMs to do low-level computations. 
"""

import openai 
import re
import math  

from tenacity import (
    retry,

yuchenlin

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers import StoppingCriteria, StoppingCriteriaList, LogitsProcessor, LogitsProcessorList

model_path = "./qlora-out-hkg_300B/merged/"

tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto", trust_remote_code=True)

yuchenlin

[
  {
    "session_id": "7f4abf64593c439f8b085585aeeda566",
    "model_input": "<|im_start|>user\nGive me an introduction over 200 words for ShangHai BMG Chemical Co., Ltd, a chemical company in Room 602, no 291 sikai road shanghai Shanghai,China<|im_end|>\n<|im_start|>assistant\n",
    "output": [
      "ShangHai BMG Chemical Co., Ltd. is a leading chemical company located in Room 602, No. 291 Sikai Road, Shanghai, China. Established in 2010, the company has since been dedicated to providing high-quality chemical products and services to its customers. With a strong R&D team and advanced production facilities, ShangHai BMG Chemical Co., Ltd. has become a trusted partner in the chemical industry, serving a wide range of clients across various sectors. The company's commitment to innovation, quality, and customer satisfaction has enabled it to maintain a competitive edge in the market.",
      "ShangHai BMG Chemical Co., Ltd. is a leading chemical company located in Room 602, No. 291 Sikai Road, Shanghai, Chin

yuchenlin

ZebraLogic: Benchmarking the Logical Reasoning Ability of Language Models

yuchenlin

from transformers import GPT2LMHeadModel, GPT2Tokenizer
import torch 

# Test the model before saving
def chat_with_model(model, tokenizer, instruction, max_length=50):
    prompt = f"### User: {instruction}\n\n### AI: "
    inputs = tokenizer.encode(prompt, return_tensors='pt').to(device)
    outputs = model.generate(inputs, max_length=max_length, num_return_sequences=1, top_k=50)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)