TuningAI/Llama2_13B_startup_Assistant

Model Name: Llama2_13B_startup_Assistant

Description:

Llama2_13B_startup_Assistant is a highly specialized language model fine-tuned from Meta's Llama2_13B. It has been tailored to assist with inquiries related to Algerian startups, offering valuable insights and guidance in these domains.

Base Model:

This model is based on the Meta's meta-llama/Llama-2-13b-chat-hf architecture, making it a highly capable foundation for generating human-like text responses.

Dataset :

This model was fine-tuned on a custom dataset meticulously curated with more than 200 unique examples. The dataset incorporates both manual entries and contributions from GPT3.5, GPT4, and Falcon 180B models.

Fine-tuning Techniques:

Fine-tuning was performed using QLoRA (Quantized LoRA), an extension of LoRA that introduces quantization for enhanced parameter efficiency. The model benefits from 4-bit NormalFloat (NF4) quantization and Double Quantization techniques, ensuring optimized performance.

Performance:

Llama2_13B_startup_Assistant exhibits improved performance and efficiency in addressing queries related to Algerian tax law and startups, making it a valuable resource for individuals and businesses navigating these areas.

Limitations:

• While highly specialized, this model may not cover every nuanced aspect of Algerian tax law or the startup ecosystem.
• Accuracy may vary depending on the complexity and specificity of questions.
• It may not provide legal advice, and users should seek professional consultation for critical legal matters.

Training procedure

The following bitsandbytes quantization config was used during training:

• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: float16

Framework versions

• PEFT 0.4.0

! huggingface-cli login

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

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=getattr(torch, "float16"),
    bnb_4bit_use_double_quant=False)
model = AutoModelForCausalLM.from_pretrained(
        "meta-llama/Llama-2-13b-chat-hf",
        quantization_config=bnb_config,
        device_map={"": 0})
model.config.use_cache = False
model.config.pretraining_tp = 1
model = PeftModel.from_pretrained(model, "TuningAI/Llama2_13B_startup_Assistant")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-13b-chat-hf" , trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"
system_message = "Given a user's startup-related question in English, you will generate a thoughtful answer in English."
while 1:
  input_text = input(">>>")
  logging.set_verbosity(logging.CRITICAL)
  prompt = f"[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n {input_text}. [/INST]"
  pipe = pipeline(task="text-generation", model=new_model, tokenizer=tokenizer, max_length=512)
  result = pipe(prompt)
  print(result[0]['generated_text'].replace(prompt, ''))