(ql:quickload :cl-gpt2) It automatically downloads a (124M parameter GPT-2) from Hugging Face (~500 MB) on first use.
(defparameter *gpt2* (cl-gpt2:load-model :gpt2-small)) (cl-gpt2:generate *gpt2* "The meaning of life is" :length 50) Model weights stored in ~/.cache/cl-gpt2/ . Option B: cl-transformer – General Transformer Library Supports BERT, GPT, etc. Requires manual weight download. ai generator lisp download
Place these in ~/lisp-models/ and point your Lisp code there. ;; Step 1 – install SBCL and Quicklisp ;; Step 2 – in REPL (ql:quickload :cl-gpt2) ;; Step 3 – load model (downloads weights automatically) (defparameter ai (cl-gpt2:load-model :gpt2-medium)) Requires manual weight download
wget https://www.gutenberg.org/files/100/100-0.txt -O shakespeare.txt A lightweight Markov chain generator (no neural nets,
| Model Source | Command / Link | |--------------|----------------| | | wget https://huggingface.co/gpt2/resolve/main/model.safetensors | | BERT | wget https://huggingface.co/bert-base-uncased/resolve/main/pytorch_model.bin | | CodeLlama (7B) | Request from Meta, then download .gguf from Hugging Face |
# Outside Lisp, using wget wget https://huggingface.co/gpt2/resolve/main/pytorch_model.bin Then convert to Lisp-native format using provided scripts. A lightweight Markov chain generator (no neural nets, purely statistical).
(ql:quickload :cl-llama) (cl-llama:load-model "path/to/llama-2-7b.Q4_K_M.gguf") (cl-llama:generate "Once upon a time") If a Lisp library expects local weights: