Skip to content

Latest commit

 

History

History
52 lines (43 loc) · 2.52 KB

README.md

File metadata and controls

52 lines (43 loc) · 2.52 KB
Raw

Texts AI

Generates texts using an LLM fine-tuned on previous messages.

GilfoyleAI

Getting started

  1. Run pip install -r requirements.txt
  2. Follow instructions below to initialize a model if needed
  3. Run python3 texts_ai.py --help for instructions

Setting up language models

Both GPT-type and LLaMa-type models are supported. I wrote the prompts and query.py to correspond with Vicuna, for LLaMa-type models, and a custom NanoGPT for GPT-type models. Instructions to configure both of them are below, but the code should be hackable to work with a variety of models.

Fine-tuning with NanoGPT
  1. Clone karpathy/nanoGPT
  2. Run training_messages.py from this repo
  3. Edit nanoGPT/data/shakespeare/prepare.py to use the file written by query.py rather than Shakespeare
  4. Run prepare.py
  5. Follow the instructions from NanoGPT to fine-tune or train using the train.bin and val.bin files created by prepare.py
Setting up Vicuna
  1. Run brew install bash
  2. Download llama.sh from this repo
  3. Modify llama.sh to only download 7B and then run it with homebrew's bash
  4. Convert to Tokenizer format using this script and this command:
python src/transformers/models/llama/convert_llama_weights_to_hf.py \
  --input_dir /path/to/downloaded/llama/weights \
  --model_size 7B \
  --output_dir /output/path
  1. Modify LLaMa with Vicuna weights:
python3 -m fastchat.model.apply_delta \
  --base-model-path /path/to/llama-7b \
  --target-model-path /path/to/output/vicuna-7b \
  --delta-path lmsys/vicuna-7b-delta-v1.1 \
  --low-cpu-mem
  1. Run blocks 2 and 3 of this script on the Vicuna bin, replacing LLaMa with Llama
  2. Run conversion scripts from llama.cpp to convert Vicuna pth to f16 and then q4_0

Helpful resources

License

Under the MIT License. Models may have their own licenses.