Work offline with Open source LLM
Choose a Text Generation Model with HuggingFace
If you're using the HuggingFace library for the first time to generate text, it will download a model for you. After that, it will use the downloaded model for future runs.
device_type: This is where you choose if you want to use your computer's regular power (CPU) or its faster power (GPU, if available). model_id: This is like a name for the model you want to use. It helps HuggingFace find the right model for you. model_basename (optional): If you're using a specific type of model, you can mention it here. What It Does: This function helps you set up a tool to create text using a specific model. It's like having a writing assistant that follows the rules of a particular writer.
The code works with special HuggingFace models that have names ending in "GPTQ" and also have certain words like ".no-act.order" or ".safetensors" in their name on the HuggingFace website. It is also designed to work with any HuggingFace models that have names ending in "-HF" or have a ".bin" file in their HuggingFace repository.
It gets a special type of model, either "HuggingFaceInstructEmbeddings" or "HuggingFaceEmbeddings". It loads a set of stored vectors from a previous step. It also loads a local language model (LLM) using a function called "load_model". You can choose different LLMs. It sets up a process for retrieving question and answer pairs. It then uses this setup to find answers to questions.
# for HF models
# model_id = "TheBloke/vicuna-7B-1.1-HF"
# model_basename = None
# model_id = "TheBloke/Wizard-Vicuna-7B-Uncensored-HF"
# model_id = "TheBloke/guanaco-7B-HF"
# model_id = 'NousResearch/Nous-Hermes-13b' # Requires ~ 23GB VRAM. Using STransformers
# alongside will 100% create OOM on 24GB cards.
# llm = load_model(device_type, model_id=model_id)
# for GPTQ (quantized) models
# model_id = "TheBloke/Nous-Hermes-13B-GPTQ"
# model_basename = "nous-hermes-13b-GPTQ-4bit-128g.no-act.order"
# model_id = "TheBloke/WizardLM-30B-Uncensored-GPTQ"
# model_basename = "WizardLM-30B-Uncensored-GPTQ-4bit.act-order.safetensors" # Requires
# ~21GB VRAM. Using STransformers alongside can potentially create OOM on 24GB cards.
# model_id = "TheBloke/wizardLM-7B-GPTQ"
# model_basename = "wizardLM-7B-GPTQ-4bit.compat.no-act-order.safetensors"
# model_id = "TheBloke/WizardLM-7B-uncensored-GPTQ"
# model_basename = "WizardLM-7B-uncensored-GPTQ-4bit-128g.compat.no-act-order.safetensors"
# for GGML (quantized cpu+gpu+mps) models - check if they support llama.cpp
# model_id = "TheBloke/wizard-vicuna-13B-GGML"
# model_basename = "wizard-vicuna-13B.ggmlv3.q4_0.bin"
# model_basename = "wizard-vicuna-13B.ggmlv3.q6_K.bin"
# model_basename = "wizard-vicuna-13B.ggmlv3.q2_K.bin"
# model_id = "TheBloke/orca_mini_3B-GGML"
# model_basename = "orca-mini-3b.ggmlv3.q4_0.bin"
Install anaconda environment. Anaconda3-2023.03-1-Windows-x86_64 was used.
pip install -r requirements.txt pip install https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-2.3.1/en_core_web_sm-2.3.1.tar.gz
Optional if it was not installed pip install sentencepiece pip install protobuf==3.19.0
Copy your text files to DATA directory.
Run the following cmd python ingest.py python generate_questions.py question_count=10 python generate_answers.py
.txt, .md, .py, .pdf, .csv, .docx, .doc"
For text file, ANSI encoding is default.
HTML reading feature will come soon